An integrated platform for handling confounding bias and missingness in real-world research data

Abstract

The present invention relates to an integrated platform for processing confounding bias and missingness for real-world research data, including: a data import and preprocessing module for reading and preprocessing an input data set; a missingness identification and processing module for identifying missingness in the input data set, and performing multiple interpolation on missing values ​​to obtain an interpolation data set; a confounding bias qualitative evaluation module for constructing a directed acyclic graph, screening confounding factors based on the directed acyclic graph, and obtaining confounding factors to be corrected; a confounding bias quantitative control module for balancing the distribution of exposure groups and control groups on confounding factors based on the input data set or interpolation data set and factors to be corrected, and obtaining a control data set; a sensitivity analysis module for evaluating the influence of uncontrolled confounding factors on the relationship between research factors and research diseases based on the control data set and outputting results. The present invention can enhance the rigor of statistical methods in real-world research.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical research data processing, and in particular to an integrated platform for processing confounding bias and missingness of real-world research data. Background Art

[0002] Real-world research refers to research conducted in a natural environment, that is, clinical research data related to patients are collected in a real-world environment, aiming to understand and evaluate the effects of medical products or interventions in actual clinical practice. Compared with clinical trials, real-world research can better reflect the situation in actual clinical practice, is more realistic, and the sample selection of the research is often more representative, and the research results are more externally valid. However, the sample selection and data collection of real-world research are conducted in a natural environment without strict control and management. Therefore, there may be problems such as low data quality, serious data missing, and difficulty in data analysis, resulting in low reliability of research results. More importantly, in a natural environment, there is an imbalance between the comparison groups in some important characteristics that affect the outcome, which makes the causal effect calculated based on the sample subject to serious confounding bias. If rigorous confounding control and data processing are not carried out, the quality of evidence in real-world research will be greatly reduced.

[0003] At present, there are many mature statistical methods for identifying and dealing with confounding bias and missing values. However, the statistical principles involved in these methods are complex. The use of statistical software such as SAS, R, and Stata requires a certain statistical background and programming skills. Influenced by their professional background, many researchers have weak statistical and programming foundations and lack systematic knowledge of identifying and dealing with confounding factors. This makes a considerable number of real-world research results face a greater risk of bias. There is an urgent need for a real-world data integration and processing platform that does not require programming and has a user-friendly interface. Summary of the invention

[0004] The purpose of the present invention is to provide an integrated platform for processing confounding bias and missingness of real-world research data, with the aim of solving the statistical application problems faced by non-statistical researchers in real-world research, enhancing the rigor of statistical methods in real-world research, and more efficiently utilizing samples in the real world to provide strong support for medical decision-making.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] An integrated platform for handling confounding bias and missingness for real-world research data, including:

[0007] A data import and preprocessing module, used to read and preprocess input data sets, wherein the input data is clinical research medical data of real-world research;

[0008] A missing value identification and processing module is used to identify missing values ​​in the input data set and perform multiple interpolation on the missing values ​​to obtain an interpolated data set;

[0009] A confounding bias qualitative evaluation module is used to construct a directed acyclic graph, screen confounding factors based on the directed acyclic graph, and obtain confounding factors to be corrected;

[0010] A confounding bias quantitative control module, used to balance the distribution of the exposure group and the control group on the confounding factors based on the input data set or the interpolation data set and the confounding factors to be corrected, and obtain a control data set;

[0011] The sensitivity analysis module is used to evaluate the influence of uncontrolled confounding factors on the relationship between research factors and research diseases based on the control data set and output the results.

[0012] Optionally, the data import and preprocessing module includes a data reading unit, a data display unit and a variable setting unit;

[0013] The data reading unit is used to read the input data set, wherein the data set includes all research variables and non-repetitive ID numbers identifying samples;

[0014] The data display unit is used to display the variable name and the first several rows of data of the input data set;

[0015] The variable setting unit is used to obtain the variable type of each variable in the input data set, and convert the variable type in the input data set into a factor type variable and a numeric type variable.

[0016] Optionally, the missing identification and processing module includes a missing result display unit and a multiple interpolation unit;

[0017] The missing result display unit is used to calculate the missing ratio and shadow matrix of each variable in the input data set, and display them through a bar chart and a heat map respectively;

[0018] The multiple interpolation unit is used to predict the distribution of missing values ​​using an interpolation method and complete data in the input data set to obtain interpolation data, and perform multiple interpolation processing on the interpolation data based on the variable type to obtain the interpolation data set, wherein the interpolation methods include mean matching, classification and regression tree, and random forest interpolation.

[0019] Optionally, the confounding bias qualitative evaluation module includes a Robins-I scale unit and a confounding factor screening unit;

[0020] The Robins-I scale unit is used to evaluate the rigor of the confounding bias control process using the Robins-I scale;

[0021] The confounding factor screening unit is used to screen the confounding factors using a directed acyclic graph, wherein the directed acyclic graph includes a saturated directed acyclic graph and a final directed acyclic graph.

[0022] Optionally, screening the confounding factors using a directed acyclic graph includes:

[0023] Acquire information for constructing the saturated directed acyclic graph, and acquire a directed edge index table based on the information, wherein the information of the saturated directed acyclic graph includes: all variables in the directed acyclic graph, exposure variables, outcome variables, variables occurring before exposure, variables occurring between exposure and outcome, and variables occurring after outcome;

[0024] Based on the directed edge index table, the forceNetwork function in the networkD3 package is used to construct the saturated directed acyclic graph, wherein there is a causal relationship between any two variable nodes in the saturated directed acyclic graph;

[0025] Based on the saturated directed acyclic graph, obtaining a directed evaluation table, and processing the directed evaluation table;

[0026] Based on the processed directed evaluation table, the final directed acyclic graph is drawn using the forceNetwork function in the networkD3 package, and the adjustment variable set is generated using the adjustmentSets function in the dagitty package. Based on the adjustment variable set, the confounding factors to be corrected are obtained.

[0027] Optionally, based on the saturated directed acyclic graph, obtaining a directed evaluation table includes:

[0028] The directed edge index table of the saturated directed acyclic graph is obtained, and the directed evaluation table is obtained by adding blank columns to the directed edge index table.

[0029] Optionally, the confounding bias quantitative control module includes a propensity score model construction unit, a propensity score distribution unit, a balance confounding factor unit and a balance test unit;

[0030] The propensity score model construction unit is used to construct a Logistic regression model using the analysis data set based on the exposure factors under study and the confounding factors to be corrected, and obtain a propensity score data set based on the Logistic regression model, wherein the analysis data set includes a complete data set and the interpolation data set, and the complete data set is a data set obtained by deleting missing samples in the input data set;

[0031] The propensity score distribution unit is used to obtain a frequency density histogram of the propensity scores of the exposure group and the control group in the propensity score data set, evaluate the overlap of the propensity score distribution of the exposure group and the control group based on the frequency density histogram, and evaluate the propensity score method according to the overlap;

[0032] The balanced confounding factor unit is used to balance the distribution of the exposure group and the control group on the confounding factors by using propensity score matching and propensity score weighting to obtain a control data set;

[0033] The balance test unit is used to calculate the descriptive statistics, difference test results and absolute standard deviation of the exposure group and the control group in the pre- and post-control data sets, wherein the pre-control data set is a propensity score data set and the post-control data set is a control data set.

[0034] Optionally, obtain the propensity score dataset including:

[0035] Based on the Logistic regression model, a list containing the fitting information of the Logistic regression model is obtained, the probability value of each sample in the list being assigned to the exposure group is calculated to generate a propensity score vector, and the propensity score vector is added as a new column to the complete data set to generate the propensity score data set.

[0036] Optionally, based on the control data set, the influence of uncontrolled confounding factors on the relationship between the research factors and the research diseases is evaluated and the output results include:

[0037] The causal effect value and the credible interval are calculated based on the control data set, and the E value is calculated based on the type of the causal effect and the credible interval. The E value is used to obtain the degree of influence of uncontrolled confounding factors on the relationship between the research factors and the research disease under the condition of controlling the measured confounding factors.

[0038] The beneficial effects of the present invention are:

[0039] The platform constructed by the present invention can correctly process real-world data, identify and handle confounding bias and missing data without understanding statistical principles when dealing with missing and confounding bias.

[0040] The present invention can, to a large extent, solve the statistical application problems faced by non-statistical researchers in real-world research, enhance the rigor of statistical methods in real-world research, more efficiently utilize real-world samples, and provide strong support for medical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0042] Figure 1 It is a structural diagram of an integrated platform for processing confounding bias and missingness of real-world research data according to an embodiment of the present invention;

[0043] Figure 2 A flowchart of an integrated platform for processing confounding bias and missingness for real-world research data according to an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of an interactive page for importing data and setting variable types in an integrated platform according to an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of an interactive page for identifying variable missing ratios in an integrated platform according to an embodiment of the present invention;

[0046] Figure 5 A schematic diagram of an interactive page for variable missing pattern recognition in an integrated platform according to an embodiment of the present invention;

[0047] Figure 6 A schematic diagram of an interactive page for multiple interpolation of missing values ​​in an integrated platform according to an embodiment of the present invention;

[0048] Figure 7 A schematic diagram of the Robins-I scale interaction page in the integrated platform of an embodiment of the present invention;

[0049] Figure 8 A schematic diagram of a saturated directed acyclic graph interaction page in an integrated platform according to an embodiment of the present invention;

[0050] Fig. 9 It is a schematic diagram of the final directed acyclic graph interaction page in the integrated platform of an embodiment of the present invention;

[0051] Fig.10 A schematic diagram of an interactive page for constructing a propensity score model in an integrated platform according to an embodiment of the present invention;

[0052] Fig.11A schematic diagram of an interactive page of a propensity score distribution graph in an integrated platform according to an embodiment of the present invention;

[0053] Fig.12 This is a schematic diagram of an interactive page for controlling confounding (taking nearest distance matching as an example) using a propensity score method in an integrated platform according to an embodiment of the present invention;

[0054] Fig.13 A schematic diagram of a balance check interaction page in an integrated platform according to an embodiment of the present invention;

[0055] Fig.14 A schematic diagram of a sensitivity analysis interaction page in an integrated platform according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] like Figure 1-2 As shown, this embodiment provides an integrated platform for processing confounding bias and missingness for real-world research data, including:

[0059] Data import and preprocessing module, used to read and preprocess input data sets;

[0060] Specifically, the data import and preprocessing module is used to read, display and set the variable types in the input data set, where the input data set is clinical research medical data from real-world research. It includes three parts: data reading, data display and setting variable types. Among them, the file format of the input data set in the data reading part is .csv format, the data format is wide data, the first line of the file is the variable name, the first column is the non-repetitive ID number that identifies the sample, and the data set contains all research variables. The data display unit part displays the variable name and the first 10 rows of data of the input data set on the first web page, such as Figure 3As shown. The data import part is composed of a "data import" input control and a "data import" function server. The "data import" input control is generated by the fileInput function in the shiny package, and is used to input the path information of the dataset file and display it on the first web page. The intervention / exposure variables of interest are binary variables. The "data import" function server is implemented by the vroom function of the vroom package in the R language. The vroom function takes the dataset file path information passed by the input control as a parameter and imports the dataset into the R language environment. The data display part is composed of a "display data" output control and a "display data" function server. The "display data" output control is generated by the tableOutput function in the shiny package and displayed on the first web page. The "display data" function server is implemented by the head function in the R language.

[0061] The variable setting unit refers to collecting the variable names of the numerical variables, categorical variables, and hierarchical variables input by the user through the input control, converting the corresponding categorical variables and hierarchical variables in the input data set into variables of factor type, and converting the corresponding numerical variables in the input data set into variables of numeric type. The variable type setting part is composed of the "set variable type" input control and the "set variable type" function server. The variable type setting input control is generated by the selectInput function in the shiny package, including the "select numerical variable" input control, the "select categorical variable" input control, and the "select hierarchical variable" input control, and is displayed in the first web page. After the data import is completed, the updateSelectInput function in the shiny package is used to update the input control options to the column name of the imported data set, which is used to set the variable type of each variable in the input data set. The "set variable type" function server uses the R language lapply function, as.numeric function, as.character function, and as.factor function to batch set the variables entered by the "set variable type" input control to the corresponding variable type.

[0062] A missing value identification and processing module is used to identify missing values ​​in the input data set and perform multiple interpolation on the missing values ​​to obtain an interpolated data set;

[0063] Specifically, the missing identification and processing module includes a missing result display unit and a multiple interpolation unit;

[0064] The missing result display unit is used to calculate the missing ratio of each variable in the input data set and draw a bar chart for display. In the R language program, it can automatically identify which variables are missing, and use the plotOutput function in the shiny package to display the bar chart on the second web page, such as Figure 4 As shown; calculate the shadow matrix of the missing variables in the input data set and draw a heat map for display. Use the md.pattern function of the mice package in R language to draw the missing pattern diagram of the selected variable set, and use the plotOutput function in the shiny package to display the diagram on a third web page, as shown Figure 5 As shown. The missing proportion part is composed of the "missing proportion" output control and the "missing proportion" function server. The missing proportion output control is generated by the plotOutput function in the shiny package, and the missing proportion function server is implemented by the plotmissing function of the DataExplorer package in the R language. The missing pattern part is composed of the "missing pattern" output control and the "missing pattern" function server. The missing pattern output control is generated by the plotOutput function in the shiny package and displayed in the third web page, and the missing pattern function server is implemented by the md.pattern function of the mice package in the R language.

[0065] The multiple interpolation unit is used to adopt the interpolation method set by the user and use the complete data in the input data set to predict the distribution of missing values. Based on this distribution, the interpolation values ​​specified by the user are generated for the number of times. Then, the variable type information set by the user in the data import and preprocessing module is used to determine the variable type of each missing variable. The mean, mode and median of the multiple interpolation values ​​of the numerical, categorical and rank missing variables are taken as the replacement values ​​of the missing values. Finally, the interpolation data set is obtained and displayed on the fourth web page through the tableOutput function. The user can download the complete data set after filling by clicking the downloadButton control, such as Figure 6 The user-defined imputation methods include "Predictive mean matching (PMM)", "Classification and regression trees (CART)" and "Random forest imputation (RF)".

[0066] The multiple interpolation part consists of the "multiple interpolation parameters" input control and the "multiple interpolation" function server. The "multiple interpolation parameters" input control is generated by the numericInput function and the selectInput function in the shiny package and displayed on the fourth web page, including the "interpolation times" input control and the "interpolation method" input control. The "multiple interpolation" function server is implemented by the mice function in mice in the R language. The mice function calculates the filling value of each missing data through the parameters passed by the "multiple interpolation parameters" input control, and fills it in the original data set to form a complete data set.

[0067] A confounding bias qualitative evaluation module is used to construct a directed acyclic graph, screen confounding factors based on the directed acyclic graph, and obtain confounding factors to be corrected;

[0068] Specifically, the confounding bias qualitative assessment module includes the Robins-I scale unit and the confounding factor screening unit;

[0069] The Robins-I scale unit is used to display the Robins-I scale for filling in, which is used by users to evaluate the rigor of the confounding bias control process and understand the standard process of confounding control, such as Figure 7 The Robins-I scale part is composed of the "Robins-I scale" input control, which is generated by the selectInput function in the shiny package and displayed on the fifth web page.

[0070] The confounding factor screening unit is used to screen confounding factors using a directed acyclic graph;

[0071] Specifically, the confounding factor screening part screens the confounding factors by drawing a directed acyclic graph, including drawing a saturated directed acyclic graph part and drawing a final directed acyclic graph part.

[0072] The saturated directed acyclic graph is the precursor of the final directed acyclic graph, that is, the final directed acyclic graph is obtained after evaluating and modifying the saturated directed acyclic graph, and then the confounding factors that need to be corrected are obtained based on the final directed acyclic graph, that is, the adjustment variable set is obtained. The final directed acyclic graph is obtained after evaluating and modifying the saturated directed acyclic graph, that is, a new directed edge index table (corresponding to the information of the final directed acyclic graph) is obtained by modifying the evaluation directed edge index table (corresponding to the information of the saturated directed acyclic graph), and then a new directed acyclic graph is generated based on the new directed edge index table, that is, the final directed acyclic graph.

[0073] The way to obtain the adjustment variable set based on the final directed acyclic graph is: based on the information in the source and target columns of the evaluated directed evaluation table, a string vector is generated. Each element in this vector uses a special format to represent the parent node of each node in the final directed acyclic graph. The format is: the name of the node ~ the name of each parent node is connected by +, for example, "y~a+b+c" is used to represent that the parent nodes of node y are a, b, and c. Then, the as.formula function is used for each element in this string vector to convert the string vector into a formula vector. Based on the formula vector and the exposure variable name and the outcome variable name obtained in the input control, the adjustment variable set is calculated using the adjustmentSets function. The principle of the adjustmentSets function is the backdoor criterion. The specific content of the backdoor criterion is: given a pair of ordered variables (X, Y) in a directed acyclic graph, if the variable set Z satisfies: there is no descendant node of X in Z, and Z blocks every path between X and Y that points to X, then Z is said to satisfy the backdoor criterion about (X, Y). If the variable set Z satisfies the backdoor criterion of (X, Y), then the causal effect of X on Y can be calculated by adjusting the variable set Z.

[0074] The part of drawing a saturated directed acyclic graph includes: using the information required to construct a saturated directed acyclic graph input by the control to generate a directed edge index table, and then using the forceNetwork function in the networkD3 package to draw a saturated directed acyclic graph based on the information in the table, such as Figure 8 As shown, it is displayed on the sixth web page. The information required to construct a directed acyclic graph includes: all variables in the directed acyclic graph, exposure variables, outcome variables, variables occurring before exposure, variables occurring between exposure and outcome, and variables occurring after outcome. The directed edge index table contains two columns: source and target, which represent the starting point and end point of the directed edge, respectively, and are used to identify different directed edges. The saturated directed acyclic graph means that there are directed edges between any two variable nodes in the directed acyclic graph, that is, there is a causal relationship between any two variable nodes in the graph.

[0075] After the saturated directed acyclic graph is generated, the saturated directed acyclic graph and the directed edge evaluation table are output through the control. The directed evaluation table adds three blank columns after the directed index table, namely temporality, face_validity, recourse_to_theory, and decision. The directed edge evaluation table is used by users to filter the directed edges in the directed acyclic graph. The table contains five columns: source, target, temporality, face_validity, recourse_to_theory, and decision. The source and target columns represent the starting point and end point of the directed edge, respectively, and are used to identify different directed edges. The temporality, face_validity, and recourse_to_theory columns contain "yes" and "no" options, which are used by users to judge whether the directed edges meet the three criteria of "temporality", "face validity", and "theoretical support". The "temporality" standard means that the starting point of the directed edge must occur before the end point. The "face validity" standard is the biological plausibility in the Hill criterion, which refers to whether the causal relationship represented by the directed edge is contrary to the facts, knowledge and theories related to the disease. The "theoretical support" standard refers to whether the causal relationship represented by the directed edge has formal theoretical support. The decision column contains "keep", "delete", and "reverse" options, which are used by users to decide whether to keep the directed edge or reverse the directed edge based on the judgment results of the three standards. Among the three standards, temporality and face validity are necessary conditions. If they are not met, the edge will be excluded. However, the lack of theoretical support alone is not enough to exclude the directed edge.

[0076] The final directed acyclic graph drawing part includes: using the control to input the user's completed directed edge evaluation table, deleting the rows with "delete" in the decision column of the table, swapping the source and target columns of the rows with "reverse" in the table, and not processing the rows with "retain" in the table. Based on the information in the source and target columns in the processed table, the forceNetwork function in the networkD3 package is used to draw the final directed acyclic graph, such as Fig. 9 As shown, shown on the sixth web page, the adjustment variable set is generated using the adjustmentSets function in the dagitty package.

[0077] A confounding bias quantitative control module, used to balance the distribution of the exposure group and the control group on the confounding factors based on the input data set or the interpolation data set and the factors to be corrected, so as to obtain a control data set;

[0078] Specifically, the confounding bias quantitative control module includes a propensity score model construction unit, a propensity score distribution unit, a balance confounder unit, and a balance test unit;

[0079] The propensity score model construction unit is used to set the model parameters through the input controls, and then use the analysis data set to build a Logistic regression model, and then use the model to calculate the probability value of each sample being assigned to the exposure group, that is, the propensity score, and merge it into the analysis data set as a new column to finally generate a propensity score data set. The analysis data set is a data set set by the user for quantitative control of confounding bias, including a complete data set and an interpolation data set, wherein the complete data set is a data set obtained after deleting the missing samples in the input data set. After obtaining the complete data set, the intervention / exposure of interest, the outcome of interest, and the confounding factors that need to be controlled are set through the three selectInput input controls. After the setting is completed, this integrated platform will use the intervention / exposure as the dependent variable and the confounding factors to be corrected as the independent variable, and use the glm function in the R language to construct a Logistc regression, generate a list containing model fitting information, and output the results of the Logistic regression in the seventh web page through the verbatimTextOutput function in the shiny package, such as Fig.10 The model parameters include the exposure factors studied and the confounders that need to be corrected, wherein the confounders that need to be corrected are set according to the adjustment variable set of the directed acyclic graph.

[0080] The part of constructing the propensity score model consists of the "variable selection" input control and the "Logistic regression" function server. The "variable selection" input control consists of the "select the dataset to be used" input control, the "select the intervention / exposure you are interested in" input control, the "select the outcome of interest" input control, and the "select the confounders to be corrected" input control, and is displayed on the seventh web page. All input controls in this part are generated by the selectInput function in the shiny package. The updateSelectInput function in the shiny package is used to update the options of the "select the intervention / exposure of interest" input control, the "select the outcome of interest" input control, and the "select the confounders to be corrected" input control to the column name of the dataset. The "Logistic regression" function server uses the information passed by the "variable selection" input control, takes the exposure / intervention of interest as the dependent variable, and the confounder as the independent variable, and uses the glm function in the R language to construct a logistic regression model, generate a list containing model fitting information, extract the conditional probability value of each sample in the list to generate a propensity score vector, and add it as a new column to the complete dataset to generate a propensity score dataset.

[0081] The propensity score distribution unit is used to draw the frequency density histograms of the propensity scores of the exposure group and the control group in the propensity score data set. This part is used to evaluate the overlap of the propensity score distribution of the exposure group and the control group. If the overlap is not good, it is not suitable to use the propensity score method to control confounding factors. For example, the balance of the data is still poor after the confounding is controlled, and a large amount of sample size will be lost after the propensity score matching. That is, the accuracy of the propensity score method is evaluated by overlap. Use the ggplot2 package in the R language to draw the propensity score distribution map in each intervention / exposure group, and use the plotOutput function in the shiny package to display the distribution map on the seventh web page, such as Fig.11 .

[0082] The confounding factor balance unit is used to set the confounding control method and corresponding parameters through the input control, balance the distribution of confounding factors in the exposure group and the control group in the analysis data set, and generate a controlled data set. The confounding control method includes propensity score matching and propensity score weighting. The confounding controlled data set is displayed in the eighth web page through the tableOutput function, and the controlled data set can be downloaded by clicking the downloadButton. The user can use the data set to calculate the causal effect value. Fig.12 .

[0083] The propensity score distribution map part consists of the "propensity score distribution map" function server and the "propensity score distribution map" output control. The "propensity score distribution map" function server uses the propensity score data in the propensity score data set and uses the ggplot2 package in the R language to draw the propensity score distribution map in each intervention / exposure group. The "propensity score distribution map" output control uses the distribution map information transmitted by the "propensity score distribution map" function server, generated by the plotOutput function in the shiny package, and displayed in the seventh web page.

[0084] The propensity score matching method is to pair samples with similar propensity scores in the exposure group and the control group, remove samples that cannot be matched, and generate a controlled data set in which the propensity scores of the exposure group and the control group have similar distributions, that is, have similar distributions of confounding factors. The parameters that need to be set in this method include the estimated causal effect type and matching method, wherein the estimated causal effect type includes the average treatment effect on the treated group (ATT) and the average treatment effect for the control group (ATC), and the matching methods include "closest distance matching", "optimal matching", "complete matching" and "genetic matching". When "closest distance matching" or "genetic matching" is selected, it is necessary to set two additional parameters, the matching ratio and the caliper value. When "optimal matching" is selected, the matching ratio parameter needs to be set additionally, and when the complete matching is selected, the caliper value parameter needs to be set additionally. The "propensity score method" input control of the confounding bias control part includes the "method selection" input control and the "parameter setting" input control, and is displayed in the eighth web page. The "method selection" input control is generated by the selectInput function in the shiny package, and the options include "propensity score matching" and "propensity score weighting". The composition of the "parameter setting" input control depends on the options of the "method selection" input control. When "propensity score matching" is selected, the "parameter setting" input control consists of the "effect type" input control, the "matching method" input control, the "matching ratio" input control, and the "caliper value" input control. The "matching ratio" input control and the "caliper value" input control are generated by the numericInput function in the shiny package, the "effect type" input control and the "matching method" input control are generated by the selectInput function in the shiny package, and all components of the "parameter setting" input control are merged into one control combination by the tabPanel function.

[0085] The propensity score weighting method is to calculate the weight of each sample in the input data set according to different weighting methods based on the propensity score of each sample in the input data set, and use it as a new column of the input data set to generate a controlled data set. This method simulates a pseudo population with similar propensity score distributions in the exposure group and the control group by weighting, that is, the exposure group and the control group in this population have similar distributions of confounding factors. When selecting propensity score weighting, the parameters that need to be set are the methods for calculating weights, including "inverse probability weighting", "stable probability weighting", and "overlap weighting". The inverse probability weighting method for calculating weights is: the weight of the individual in the exposure / intervention group Wt = 1 / PS, and the weight of the individual in the control group Wc = 1 / (1-PS), where PS is the propensity score of the individual. The stable probability weighting method for calculating weights is: the weight of the individual in the exposure / intervention group Wt = Pt / PS, and the weight of the individual in the control group Wc = (1-Pt) / (1-PS), where Pt is the proportion of the exposure / intervention group. The weights of overlapping weighting are calculated as follows: the weight of individuals in the exposure / intervention group Wt=1-PS, and the weight of individuals in the control group Wc=PS.

[0086] The balance test unit is used to calculate the descriptive statistics, difference test results and absolute standard deviation (Absolute standardized difference, ASD) of the exposure group and the control group in the data set before and after control, and generate the baseline table before and after control and the absolute standard deviation to draw the dot plot. This part is used to judge the effect of confounding control. The pre-control data set is the propensity score data set. When the variable type of the confounding factor is a numerical variable, the descriptive statistics calculated are the median and the interquartile range, the difference test used is the Wilcoxon rank sum test, and the absolute standard deviation is calculated as follows:

[0087]

[0088] in, refers to the mean and standard deviation of the confounders in the exposure / intervention group, refers to the mean and standard deviation of the confounding factors in the control group.

[0089] When the variable type of the confounding factor is a categorical variable or a rank variable, the descriptive statistics calculated are frequencies, the difference test used is the exact probability method, and the absolute standard deviation is calculated as follows:

[0090]

[0091] Among them, P sc refers to the ratio of confounding factors in the exposure / intervention group, P nsc It refers to the proportion of confounding factors in the control group.

[0092] If the standardized deviation of a confounding factor is greater than 10%, it is considered that the control effect of the confounding factor is not good, and it is necessary to reselect the method or set the parameters for control.

[0093] The "balance test" function server includes the "baseline table" function server and the "absolute standardized difference ASD plot" function server. The "baseline table" function server first uses the CreateTableOne function in the tableone package in the R language to generate a baseline table for the complete data set after multiple imputation, and then uses the R language switch function to realize the generation of a baseline table after confounding control by applying different functions according to different options in the "method selection" input control. When "propensity score matching" is selected, the baseline table of the matching data set is generated by the CreateTableOne function in the tableone package in the R language. When "propensity score weighting" is selected, the baseline table of the weighted data set is generated by the svydesign function in the survey package and the svyCreateTableOne function in the tableone package. The "absolute standardized difference ASD plot" function server extracts the ASD values ​​before and after the control of confounding bias through the ExtractSmd function in the tableone package, and draws the ASD plots of each confounding factor before and after the control through the ggplot2 package. If the standardized deviation of the confounding factor is greater than 10%, it is considered that the control effect of the confounding factor is not good, and it is necessary to reselect the method or set the parameters for control. The "balance test" output control of the confounding bias control part outputs the baseline table by the tableOutput function and the ASD graph by the plotOutput function, and is displayed in the eighth web page.

[0094] After completing the confounding control using the propensity score method, this platform generates a baseline table for the complete data set through the CreateTableOne function in the tableone package in the R language as the uncontrolled baseline table. Then, the R language switch function is used to generate a baseline table after confounding control based on different propensity score methods and different functions. When using propensity score matching, the CreateTableOne function in the tableone package in the R language is used to generate a baseline table for the matching data set. When using weighted scores, the svydesign function in the survey package and the svyCreateTableOne function in the tableone package are used to generate a baseline table for the weighted data set. Then, the ExtractSmd function in the tableone package is used to extract the ASD values ​​before and after controlling the confounding bias, and the ggplot2 package is used to draw the ASD graph of each confounding factor before and after control. If the standardized deviation of the confounding factor is greater than 10%, it is considered that the control effect of the confounding factor is not good, and it is necessary to reselect the method or set the parameters for control. The above baseline table and ASD graph are generated in the eighth web page through the tableOutput function and plotOutput function, respectively. Fig.13 .

[0095] The sensitivity analysis module is used to obtain the influence of uncontrolled confounding factors on the relationship between research factors and research diseases based on the control data set and output the results.

[0096] Specifically, the evaluation of the degree of influence of uncontrolled confounding factors on the relationship between research factors and research diseases consists of the E-value calculation part. The role of the E-value is to evaluate the severity of the problems that may be caused by uncontrolled confounding factors. When the E-value is small, it means that only a small causal relationship between the uncontrolled confounding factors and the outcome can completely eliminate the causal relationship between the intervention / exposure and the outcome calculated by the researcher. At this time, the causal relationship in the research conclusion is more likely to be a false causal relationship caused by uncontrolled confounding factors. On the contrary, when the E-value is large, it means that a large causal relationship is required between the uncontrolled confounding factors and the outcome to completely eliminate the causal relationship between the intervention / exposure and the outcome calculated by the researcher. At this time, the causal relationship in the research conclusion is less likely to be a false causal relationship caused by uncontrolled confounding factors. The calculation of the causal effect value and the credible interval generates a control data set through the confounding bias quantitative control module. After the user exports this data set, he needs to calculate the causal effect value and credible interval by himself through other software, and then calculate the E-value.

[0097] The E-value calculation part includes: setting the parameters for E-value calculation through input controls, using the EValue package to calculate the E-value, and outputting the calculation results. The parameters for E-value calculation that need to be set depend on the type of causal effect. Different effect types are selected. When the causal effect is relative risk (RR), the set parameters include point estimate, lower limit of confidence interval, upper limit of confidence interval, and invalid value. When the causal effect is odds ratio (OR), the set parameters include point estimate, lower limit of confidence interval, upper limit of confidence interval, invalid value, and whether the result incidence is <15%. When the causal effect is hazard ratio (HR), the set parameters include point estimate, lower limit of confidence interval, upper limit of confidence interval, invalid value, and whether the result incidence is <15%. When the causal effect is mean difference (MD), the set parameters include point estimate, standard error, and invalid value. This platform calculates the E-value through the EValue package in the R language, and displays it in the ninth web page through the tableOutput function. As Fig.14 .

[0098] The "E-value calculation" input control in the E-value calculation section consists of the "Effect type" input control and the "Parameter selection" input control, and is displayed on the ninth web page. The "Effect type" input control is generated by the selecInput function in the shiny package, and the options include "RR", "OR", "HR", and "MD". The "Parameter selection" input control displays different controls according to the different options of the "Effect type" input control. When "RR" is selected, the "Parameter selection" input control includes the "Point estimate" input control, the "Confidence interval lower limit" input control, the "Confidence interval upper limit" input control, and the "Invalid value" input control, all of which are generated by the numericInput function in the shiny package. When "OR" is selected, the "Parameter Selection" input control includes the "Point Estimate" input control, the "Confidence Interval Lower Limit" input control, the "Confidence Interval Upper Limit" input control, the "Invalid Value" input control, and the "Is the Result Incidence <15%" input control. The "Point Estimate" input control, the "Confidence Interval Lower Limit" input control, the "Confidence Interval Upper Limit" input control, and the "Invalid Value" input control are generated by the numericInput function in the shiny package, and the "Is the Result Incidence <15%" input control is generated by the selectInput function in the shiny package, and the options include "Yes" and "No". When "HR" is selected, the "Parameter Selection" input control includes the "Point Estimate" input control, the "Confidence Interval Lower Limit" input control, the "Confidence Interval Upper Limit" input control, the "Invalid Value" input control, and the "Is the Result Incidence <15%" input control. The "Point Estimate" input control, the "Confidence Interval Lower Limit" input control, the "Confidence Interval Upper Limit" input control, and the "Invalid Value" input control are generated by the numericInput function in the shiny package, and the "Is the Result Incidence <15%" input control is generated by the selectInput function in the shiny package, and the options include "Yes" and "No". When "MD" is selected, the "Parameter Selection" input control includes the "Point Estimate" input control, the "Standard Error" input control, and the "Invalid Value" input control, all of which are generated by the numericInput function in the shiny package.

[0099] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An integrated platform for handling confounding bias and missingness for real-world research data, characterized by: include: A data import and preprocessing module, used to read and preprocess input data sets, wherein the input data is clinical research medical data of real-world research; A missing value identification and processing module is used to identify missing values ​​in the input data set and perform multiple interpolation on the missing values ​​to obtain an interpolated data set; A confounding bias qualitative evaluation module is used to construct a directed acyclic graph, screen confounding factors based on the directed acyclic graph, and obtain confounding factors to be corrected; The confounding bias qualitative evaluation module includes a Robins-I scale unit and a confounding factor screening unit; The Robins-I scale unit is used to evaluate the rigor of the confounding bias control process using the Robins-I scale; The confounding factor screening unit is used to screen the confounding factors using a directed acyclic graph, wherein the directed acyclic graph includes a saturated directed acyclic graph and a final directed acyclic graph; Using directed acyclic graphs to screen the confounding factors includes: Acquire information for constructing the saturated directed acyclic graph, and acquire a directed edge index table based on the information, wherein the information of the saturated directed acyclic graph includes: all variables in the directed acyclic graph, exposure variables, outcome variables, variables occurring before exposure, variables occurring between exposure and outcome, and variables occurring after outcome; Based on the directed edge index table, the forceNetwork function in the networkD3 package is used to construct the saturated directed acyclic graph, wherein the saturated directed acyclic graph refers to a directed acyclic graph in which there are directed edges between any two variable nodes, that is, there is a causal relationship between any two variable nodes in the graph, and the directed edge index table includes two columns: source and target, which respectively represent the starting point and the end point of the directed edge, and are used to identify different directed edges; Based on the saturated directed acyclic graph, obtaining a directed evaluation table, and processing the directed evaluation table; The directed evaluation table adds 4 blank columns after the directed index table, namely timeliness, face validity, theoretical support, and decision, which are used by users to judge whether the directed edge meets the three criteria of "timeliness", "face validity" and "theoretical support". The "timeliness" criterion means that the starting point of the directed edge must occur before the end point. The "face validity" criterion is the biological plausibility in the Hill criterion, which refers to whether the causal relationship represented by the directed edge is contrary to the facts, knowledge and theories related to the disease. The "theoretical support" criterion refers to whether the causal relationship represented by the directed edge has formal theoretical support. The decision column contains "retain", "delete", and "reverse" options, which are used by users to decide whether to keep the directed edge or reverse it according to the judgment results of the three criteria. Among the three criteria, timeliness and face validity are necessary conditions. If they are not met, the edge will be excluded. However, the lack of theoretical support alone is not enough to exclude the directed edge. Processing the directed evaluation form includes the user filling in the directed evaluation form; Based on the processed directed evaluation table, the forceNetwork function in the networkD3 package is used to draw the final directed acyclic graph, and the adjustment set of variables is generated by the adjustmentSets function in the dagitty package. Based on the adjustment set of variables, the confounding factors to be corrected are obtained; A confounding bias quantitative control module, used to balance the distribution of the exposure group and the control group on the confounding factors based on the input data set or the interpolation data set and the confounding factors to be corrected, and obtain a control data set; The confounding bias quantitative control module includes a propensity score model construction unit, a propensity score distribution unit, a balance confounding factor unit and a balance test unit; The propensity score model construction unit is used to construct a Logistic regression model using the analysis data set based on the exposure factors under study and the confounding factors to be corrected, and obtain a propensity score data set based on the Logistic regression model, wherein the analysis data set includes a complete data set and the interpolation data set, and the complete data set is a data set obtained by deleting missing samples in the input data set; The propensity score distribution unit is used to obtain a frequency density histogram of the propensity scores of the exposure group and the control group in the propensity score data set, evaluate the overlap of the propensity score distribution of the exposure group and the control group based on the frequency density histogram, and evaluate the propensity score method according to the overlap; The balanced confounding factor unit is used to balance the distribution of the exposure group and the control group on the confounding factors by using propensity score matching and propensity score weighting to obtain a control data set; The balance test unit is used to calculate the descriptive statistics, difference test results and absolute standard deviation of the exposure group and the control group in the data sets before and after control, wherein the data set before control is the propensity score data set and the data set after control is the control data set; When the variable type of the confounding factor is a numerical variable, the descriptive statistics calculated are the median and interquartile range, the difference test used is the Wilcoxon rank sum test, and the absolute standard deviation is calculated as follows: in, refers to the mean and standard deviation of the confounders in the exposure / intervention group, refers to the mean and standard deviation of the confounding factors in the control group; When the variable type of the confounding factor is a categorical variable or a rank variable, the descriptive statistics calculated are frequencies, the difference test used is the exact probability method, and the absolute standard deviation is calculated as follows: Among them, P sc refers to the ratio of confounding factors in the exposure / intervention group, P nsc It refers to the proportion of confounding factors in the control group; If the standardized deviation of the confounding factor is greater than 10%, it is considered that the control effect of the confounding factor is not good, and it is necessary to reselect the method or set the parameters for control; The sensitivity analysis module is used to evaluate the influence of uncontrolled confounding factors on the relationship between research factors and research diseases based on the control data set and output the results.

2. The integrated platform for processing confounding bias and missingness for real-world research data according to claim 1, characterized in that: The data import and preprocessing module includes a data reading unit, a data display unit and a variable setting unit; The data reading unit is used to read the input data set, wherein the data set includes all research variables and non-repetitive ID numbers identifying samples; The data display unit is used to display the variable name and the first several rows of data of the input data set; The variable setting unit is used to obtain the variable type of each variable in the input data set, and convert the variable type in the input data set into a factor type variable and a numeric type variable.

3. The integrated platform for processing confounding bias and missingness for real-world research data according to claim 2, characterized in that: The missing identification and processing module includes a missing result display unit and a multiple interpolation unit; The missing result display unit is used to calculate the missing ratio and shadow matrix of each variable in the input data set, and display them through a bar chart and a heat map respectively; The multiple interpolation unit is used to predict the distribution of missing values ​​using an interpolation method and complete data in the input data set to obtain interpolation data, and perform multiple interpolation processing on the interpolation data based on the variable type to obtain the interpolation data set, wherein the interpolation methods include mean matching, classification and regression tree, and random forest interpolation.

4. The integrated platform for processing confounding bias and missingness for real-world research data according to claim 1, characterized in that: Based on the saturated directed acyclic graph, obtaining a directed evaluation table includes: The directed edge index table of the saturated directed acyclic graph is obtained, and the directed evaluation table is obtained by adding blank columns to the directed edge index table.

5. The integrated platform for processing confounding bias and missingness for real-world research data according to claim 1, characterized in that: Obtaining the propensity score dataset includes: Based on the Logistic regression model, a list containing the fitting information of the Logistic regression model is obtained, the probability value of each sample in the list being assigned to the exposure group is calculated to generate a propensity score vector, and the propensity score vector is added as a new column to the complete data set to generate the propensity score data set.

6. The integrated platform for processing confounding bias and missingness for real-world research data according to claim 1, characterized in that: Based on the control data set, the influence of uncontrolled confounding factors on the relationship between the research factors and the research diseases is evaluated and the output results include: The causal effect value and the credible interval are calculated based on the control data set, and the E value is calculated based on the type of the causal effect and the credible interval. The E value is used to obtain the degree of influence of uncontrolled confounding factors on the relationship between the research factors and the research disease under the condition of controlling the measured confounding factors.

Images