Causal effect estimation method and system based on high-throughput medical science data variables

Abstract

The invention discloses a causal effect estimation method and system based on a high-throughput medical science data variable, and the method employs the high-throughput medical science data variable to approach to a pseudo-randomization design, so as to obtain unbiased causal effect estimation. Through optimal combination of evidence-based medical design and causal inference theory methods such as new user queue design, L1 regularization tendency score matching, negative control and synthetic positive control, P value calibration and confidence interval calibration, personalized deep causal learning algorithm and the like, a brand new intelligent big data real world research design and analysis process is invented; and an automatic or semi-automatic, universal and robust research design and analysis general solution is provided for realizing evidence-based evidence of high-level observation research by using big data.

Description

Technical Field

[0001] The present invention relates to the technical field of evidence - based medical reasoning, and particularly to a method and system for estimating causal effects based on high - throughput medical science data variables. Background Art

[0002] The statements in this section merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] With the wide popularization of Electronic Medical Record (EMR), Electronic Health Records (EHR) and high - throughput biomics, etc., it provides an opportunity for medical researchers and practitioners to obtain a large amount of covariate information. Whether in experimental studies or observational studies, in addition to obtaining conventional baseline, treatment / intervention and outcome variable information, it is easy to obtain a large amount of covariate information such as various omics markers, clinical indicators, laboratory indicators, imaging markers, etc. For example, big - data randomized clinical trials (BRCT) containing genomics and other information have been put on the agenda; in large - scale observational studies, high - throughput whole - genome biomarkers such as SNPs (Single Nucleotide Polymorphisms) have become standard for cohort baselines; similarly, similar covariate information in clinical practice can also be obtained from large databases such as EMR / EHR / LIS (Hospital Laboratory Information System). In this context, all stakeholders (such as the pharmaceutical industry, regulatory agencies, Health Technology Assessment (HTA) agencies, etc.) have begun to use the data generated in clinical practice as a supplement to randomized controlled trial data in order to obtain more generalizable evidence - based medical evidence.

[0004] As a representative of experimental causal evidence acquisition, the Randomized Controlled Trial (RCT) effectively avoids numerous confounding biases through "artificial randomization" in the design phase and is a standard of evidence-based medicine with internal validity. However, in order to pursue ideal internal validity, RCT often adopts very strict inclusion and exclusion criteria and does not randomly select trial subjects, thus inevitably resulting in poor external validity due to selection bias. Currently, a large number of studies have shown that the trial samples of many RCTs have poor representativeness of the target population, and the external validity and generalization ability of their causal evidence are not good.

[0005] In addition, some studies have combined big data and Real World Study (RWS), and achieved quasi-randomization through methods such as generalized regularized propensity score matching. Although the external validity of causal evidence has been improved, due to the inability to identify the temporal relationship between covariates and intervention / exposure factors and not meeting the condition that the time precedence relationship serves as a causal inference condition, it is very easy to mis-match mediator variables and collider variables, thus resulting in a biased estimate of the causal effect. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a method and system for estimating causal effects based on high-throughput medical science data variables, which uses high-throughput medical science data variables to approximate quasi-randomized design to obtain an unbiased estimate of the causal effect.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect, the present invention provides a method for estimating causal effects based on high-throughput medical science data variables, including:

[0009] Designing a new user cohort based on high-throughput medical science data, and configuring the cohort according to the analysis needs and configuration rules of the group or individual for the disease to be analyzed, where the new user cohort includes exposure cohorts and outcome cohorts of the target group and the control group;

[0010] Configuring a high-throughput covariate set, fitting an L1-regularized regression model of the configured exposure cohort and the high-throughput covariate set to estimate the propensity score and quasi-randomly process the configured new user cohort, and when the quasi-randomization verification is passed, obtaining a quasi-randomized data set of the cohort to be analyzed;

[0011] For the quasi-randomized dataset of the cohort to be analyzed, negative control validation is performed using exposure-outcome pairs known to have no causal effect, and synthetic positive control validation is performed using exposure-outcome pairs known to have causal effects and effect estimates, to verify the existence of unobserved residual confounding variables and estimation errors;

[0012] When there are residual confounding variables and estimation errors, P-value calibration and confidence interval calibration are performed on the selected outcome prediction model. After calibration, according to the quasi-randomized dataset of the cohort to be analyzed, the average causal effect estimate or individual causal effect estimate of the population is obtained using the outcome prediction model.

[0013] As an alternative implementation, the process of designing a new user cohort and performing cohort configuration includes: establishing a medical science data warehouse based on high-throughput medical science data and defining a concept set, and extracting data for cohort configuration according to set standards and configuration rules;

[0014] Among them, the standards include: domain, that is, the source of data stored in the medical science data warehouse; concept set, that is, the composition of concepts of clinical features, and different concepts come from the same or different domains; temporal logic, that is, the time relationship between the included concepts and the occurrence of events; the configuration rules include inclusion events, inclusion criteria, and exclusion criteria.

[0015] As an alternative implementation, the process of obtaining the quasi-randomized dataset of the cohort to be analyzed includes: setting all covariates other than the outcome and exposure as the covariate set, and performing quasi-randomization verification using overlapping-preference score evaluation and balance diagnosis.

[0016] As an alternative implementation, negative control validation includes: selecting a variable that has no causal effect relationship with the outcome as the negative control exposure, or selecting a variable that has no causal effect relationship with the exposure as the negative control outcome. After adjusting the known observable confounding variables, if there is still a causal effect relationship between the exposure and the outcome, it indicates that there are still residual confounding variables.

[0017] As an alternative implementation, synthetic positive control validation includes:

[0018] By adding an additional number of simulated outcomes to the negative control outcome to transform the negative control outcome into a positive control outcome; among them, a prediction model for the negative control outcome is constructed using the high-throughput covariate set, individual prediction probabilities are obtained based on the prediction model, and according to the set true effect value, the number of outcomes to be increased is determined. Thus, the corresponding number of negative control outcomes is changed to synthetic positive control outcomes according to the individual prediction probabilities. The synthetic positive control outcomes include the true negative control outcomes and simulated outcomes, and maintain the same bias source as the original negative control outcomes and covariates;

[0019] For a negative control variable that has no causal effect relationship with the outcome or exposure, after artificially adding a causal effect, calculate the simulated incidence rate of the outcome event; compare the difference between the true outcome incidence rate and the simulated outcome event incidence rate to determine whether there are residual confounding factors or estimation errors.

[0020] As an alternative implementation, P-value calibration includes: deriving an empirical null distribution from the actual effect estimate of the negative control, and recalculating the calibrated P-value based on this to obtain the P-value under unbiased estimation;

[0021] Confidence interval calibration includes: achieving confidence interval calibration based on the negative control outcome and the synthetic positive control outcome; specifically, fitting a systematic error model using the effect estimates of the positive and negative controls, and using the systematic error model to calculate the calibrated point estimate and confidence interval corresponding to the actual effect estimate.

[0022] As an alternative implementation, the estimation process of the average causal effect estimate of the population includes: using the Cox proportional hazards regression model as the selected outcome prediction model, and obtaining the average causal effect estimate of the population after testing the assumption conditions of the outcome prediction model;

[0023] The estimation process of the individual causal effect estimate includes: estimating the individual causal effect estimate based on counterfactual potential outcomes learning.

[0024] In a second aspect, the present invention provides a causal effect estimation system based on high-throughput medical science data variables, including:

[0025] A cohort construction module, configured to design a new user cohort based on high-throughput medical science data, and configure the cohort according to the analysis requirements and configuration rules of the population or individual for the disease to be analyzed. The new user cohort includes exposure cohorts and outcome cohorts of the target group and the control group;

[0026] A quasi-randomization module, configured to configure a high-throughput covariate set, fit an L1-regularized regression model of the configured exposure cohort and the high-throughput covariate set to estimate the propensity score and perform quasi-randomization processing on the configured new user cohort. When the quasi-randomization is verified, a quasi-randomized dataset to be analyzed is obtained;

[0027] A verification module, configured to perform negative control verification on the quasi-randomized dataset to be analyzed using exposure-outcome pairs known to have no causal effect, and perform synthetic positive control verification using exposure-outcome pairs known to have causal effects and effect estimates to verify whether there are unobserved residual confounding variables and estimation errors;

[0028] The causal effect estimation module is configured to calibrate the P-value and confidence interval of the selected outcome prediction model when there are residual confounding variables and estimation errors. After the calibration is completed, according to the quasi-randomized dataset of the cohort to be analyzed, the outcome prediction model is used to obtain the estimated value of the average causal effect or the estimated value of the individual causal effect of the population.

[0029] In a third aspect, the present invention provides an electronic device, including a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the method described in the first aspect is completed.

[0030] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the method described in the first aspect is completed.

[0031] In a fifth aspect, the present invention provides a computer program product, including a computer program. When the computer program is executed by a processor, the method described in the first aspect is implemented.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] The present invention provides a method for estimating causal effects based on high-throughput medical science data variables. Based on big data real-world research, it uses high-throughput medical science data variables to approximate quasi-randomized design to obtain unbiased causal effect estimates. This method optimally combines evidence-based medical design and causal inference theoretical methods such as new user cohort design, L1-regularized propensity score matching, negative control and synthetic positive control, P-value calibration and confidence interval calibration, and personalized deep causal learning algorithm, and invents a brand-new intelligent big data real-world research design and analysis process, providing an automated or semi-automated, generalized, and robust research design and analysis general solution for using big data to achieve evidence-based evidence for high-level observational studies.

[0034] Based on a large number of covariates, the present invention transforms the new user cohort design to meet the causal inference conditions of the time precedence relationship. While improving the external validity and generalization extrapolation ability of causal evidence, through time factors, it avoids mis-matching or adjusting mediator variables and collider variables, and obtains approximately unbiased evidence-based causal effect estimates.

[0035] The present invention completes the matching or weighting between comparison groups through regularized propensity scores of large-scale covariates in the big data cohort warehouse. At the same time, in order to ensure the quality of quasi-randomization in this process, quasi-randomization verification is carried out through overlap-preference score evaluation and balance diagnosis. When the quasi-randomization verification is passed, it indicates that effective quasi-randomization processing has been completed on the new user cohort data.

[0036] After regularized propensity score matching and diagnosis, the bias caused by large-scale observable confounding variables has been controlled. However, there may still be some unobserved confounding variables, resulting in unknown bias in the causal effect estimation, and systematic estimation errors such as selection bias and measurement error cannot be controlled by PS. Therefore, the present invention adopts negative control and synthetic positive control design strategies in causal effect estimation diagnosis to diagnose whether there are unobserved confounding bias and systematic errors.

[0037] In addition to being used to evaluate the bias of the effect estimate after the propensity score, the negative control and the positive control also need to be further empirically calibrated for the P value, point estimate and confidence interval estimated by the outcome model when possible systematic errors or biases are detected. Therefore, the present invention adopts P value calibration and confidence interval calibration to calibrate residual confounding and other systematic errors. After regularized PS, P value and confidence interval calibration, the research results can be regarded as causal estimates after controlling known confounding and unknown biases as much as possible, and can be recommended as evidence in evidence-based medicine.

[0038] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0040] Figure 1 A flowchart of a causal effect estimation method based on high-throughput medical science data variables provided in Example 1 of the present invention;

[0041] Figure 2 A schematic diagram of inclusion and exclusion criteria for constructing a new user cohort using diabetes as an example provided in Example 1 of the present invention;

[0042] Figure 3 A schematic diagram of a new user queue design using diabetes as an example provided in Example 1 of the present invention;

[0043] Figure 4 The specific structure of the new user queue provided in Example 1 of the present invention;

[0044] Figure 5 A propensity score calculation model diagram provided in Example 1 of the present invention;

[0045] Figure 6 This is a propensity score diagnosis example provided in Example 1 of the present invention;

[0046] Figure 7 Schematic diagram of synthesizing a positive control from a negative control provided in Embodiment 1 of the present invention;

[0047] Figure 8 Schematic diagram of the individual causal effect estimation process using the Uplift model provided in Embodiment 1 of the present invention. Among them, (a) takes the deep causal neural network algorithm as an example, (b) shows the values of Y when the prediction samples are in the treatment group and the control group, (c) is the uplift histogram; (d) is the ROC curve; (e) is the histogram of the PT-PC difference for each sample predicted one by one in the treatment group and the control group; (f) is the subgroup identification result using a decision tree based on the PT-PC difference of each sample. Detailed implementation manners

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0050] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that the terms "comprising" and "including" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those clearly listed steps or units, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0051] In the case of no conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other.

[0052] Embodiment 1

[0053] This embodiment provides a method for estimating causal effects based on high-throughput medical science data variables. Based on big data real world research (BRWS), high-throughput medical science data variables are used to approximate quasi-randomized design (simulating RCT) to obtain unbiased causal effect estimation.

[0054] This method relies on 140,000 source data covariates in 7 source data domains including diseases, surgeries, drugs, tests, and clinical data covered in the constructed Smart Medical Science Big Data Warehouse (SMART_MSDW). Through the optimized combination of evidence-based medical design and causal inference theoretical methods such as new user cohort design, L1-regularized propensity score matching, negative control and synthetic positive control, P-value calibration and confidence interval calibration, and personalized deep causal learning algorithm, a brand-new intelligent big data real-world research (SMART_BRWS) design and analysis process is invented, providing an automated or semi-automated, generalized, and robust research design and analysis general solution for using big data to achieve evidence-based evidence for high-level observational studies.

[0055] It mainly includes the following two aspects:

[0056] 1. Based on a large number of covariates, the transformation of the new user cohort design meets the causal inference conditions of the time precedence relationship. While improving the external validity and generalization ability of causal evidence, through time factors, it avoids mis-matching or adjusting mediating variables and collider variables, and obtains an approximately unbiased evidence-based causal effect estimate.

[0057] 2. A "big data real-world research design and analysis strategy" is proposed for big data real-world research, constructing an analysis strategy and design for generating evidence-based evidence that integrates "concept set - domain cohort warehouse - new user cohort design - high-throughput L1-regularized propensity score matching - average causal effect estimation - negative control / synthetic positive control - P-value calibration / confidence interval calibration - personalized effect estimation".

[0058] The specific process is as follows:

[0059] Design a new user cohort based on high-throughput medical science data. For the disease to be analyzed, configure the cohort according to the analysis needs and configuration rules of the group or individual. The new user cohort includes exposure cohorts and outcome cohorts for the target group and the control group;

[0060] Configure a high-throughput covariate set, fit the L1-regularized regression model of the configured exposure cohort and the high-throughput covariate set to estimate the propensity score and quasi-randomize the configured new user cohort. When the quasi-randomization is verified, obtain the quasi-randomized dataset to be analyzed for the cohort;

[0061] For the quasi-randomized dataset to be analyzed for the cohort, conduct negative control verification using exposure-outcome pairs known to have no causal effect, and conduct synthetic positive control verification using exposure-outcome pairs known to have causal effects and effect estimates to verify whether there are unobserved residual confounding variables and estimation errors;

[0062] When there are residual confounding variables and estimation errors, the selected outcome prediction model is calibrated for P value and confidence interval. After calibration, the outcome prediction model is used according to the quasi-randomized cohort data set to be analyzed to obtain the average causal effect estimate of the group or the individual causal effect estimate.

[0063] Combine the following Figure 1 The method of this embodiment is described in detail.

[0064] Step S1: Establish a causal inference framework for a new user cohort based on high-throughput medical science data;

[0065] Specifically include:

[0066] Step S11: Acquire real-world data and create a Common Data Model (CDM) standard system to provide standards for large-scale data collection, integration, mining and analysis, ensure data universality and sharing, achieve full-process data governance standardization, and transform multi-source, heterogeneous, and isolated data sources into scientific big data.

[0067] Specific data standards include the Resources Common Data Model (RCDM) and the Science-Data Common Data Model (SCDM) standards. Through extraction-transformation-load (ETL), data governance, data warehouse and other technologies, a medical science data warehouse is established to solve bottleneck problems such as poor data compliance and lack of cohort standards, providing high-quality medical science data sets for real-world evidence research.

[0068] Here are the steps:

[0069] (1) Based on the big data information such as EMR and EHR in the information systems of medical and health institutions at all levels and types, a standard code-alias mapping library for diseases, surgeries, drugs, tests, and medical terms is constructed.

[0070] For example, based on EMR / EHR data, four standard code-alias mapping libraries are created for diseases (35,862 ICD-10 codes corresponding to 6 million aliases), surgeries (12,855 ICD-CM-9 codes corresponding to 300,000 aliases), drugs (27,174 YPID codes corresponding to 1.02 million aliases), and tests (1,401 CNAS-AL09s) with massive alias information. This database provides knowledge priors for data standardization.

[0071] (2) Based on the RCDM standard, the distributed ETL technology is adopted to collect data from cross-domain, multi-source, heterogeneous data sources such as EMR and EHR, ensuring findability, accessibility, interoperability, and reusability (FAIR principle), thereby constructing the standardization work of real-world data.

[0072] (3) According to the Science-Data Common Data Model (SCDM) standard, using Natural Language Processing (NLP) as a tool, a deep learning model for medical coding standardization is constructed (such as the Roberta-Bilstm-CRF deep neural network model, etc.), realizing the real-time automatic coding and pairing standardization of "diseases, surgeries, drugs, tests, and medical terms", and constructing a medical science data warehouse; this data warehouse is designed according to the longitudinal cohort (the first column is ID, the second column is time t, and the columns after the third column are variables x1, x2,..., xp), and the total number of columns (number of variables) reaches more than 140,000 according to the SCDM standard.

[0073] (4) With the standard code-alias mapping library and the medical science data warehouse as a priori, through the integration of technologies such as blockchain, edge computing, cloud computing, federated learning, causal Meta-analysis, and secure multi-party computing, a distributed data platform supported by a complementary alliance chain of centralized and decentralized advantages is constructed to achieve multi-user concurrent access, support multi-center RWE research, and secure shared computing.

[0074] Step S12: Construct a concept set and define the standard concepts of one or more clinical features of interest. A concept set is a collection of concepts that can be reused in the analysis and is a standardized, computer-executable coding list used in real-world evidence research.

[0075] Step S13: Create a real-world research cohort. According to the specific research question hypothesis and research purpose, based on the medical science data warehouse and the constructed concept set, extract samples and their data according to specific inclusion and exclusion criteria, and construct a real-world research cohort under the principle of New User Cohort Design, which is the key to obtaining real-world evidence (RWE).

[0076] In real-world studies, the new user cohort design is similar to the clinical randomized controlled trial (RCT) design. The difference is that the former cannot adopt random allocation to eliminate the influence of confounding variables. The basic principle of the new user cohort design is as follows: observe from the time when the subjects start receiving the treatment intervention, and after a specific period of follow-up, compare the intensity differences of the outcome cohort events (such as the difference in probabilities between the two groups, the difference in measurement indicators between the two groups) between the target cohort group and the control cohort group.

[0077] This design needs to meet five elements: ① the target cohort, that is, the cohort receiving the target treatment; ② the control cohort, that is, the cohort receiving the control treatment; ③ the outcome cohort, that is, the cohort representing the target outcome population; ④ the Time-at-Risk (TAR), also known as the risk window, that is, the time period considering the outcome event; ⑤ the model, that is, after using the model to correct the differences between the target group and the control group, to evaluate the effect of the exposure factor.

[0078] Taking diabetes as an example, the inclusion and exclusion criteria for constructing the new user cohorts of insulin and metformin are as Figure 2 shown. The schematic diagrams of the new user cohorts of insulin and metformin and the cohort elements are as Figure 3 shown. The users of the two drugs are the target (exposed) cohort and the control cohort respectively, and the cohort composed of the cancer outcome occurrence population is the outcome cohort. The research analysis is to combine and analyze the three cohorts through a suitable model.

[0079] Figure 4 The following is the specific structure of the new user cohort. The start and end of the observation period represent the earliest and latest times when the research samples appear in the database respectively. The start date of the treatment plan (also known as the index date) is the start time of the cohort. The follow-up time is from the index date to the occurrence of the outcome or the end of the follow-up (the outcome has not occurred). Different from the setting of traditional epidemiological cohort studies, this design introduces the concept of the observation period, so as to ensure that all individuals have sufficient observation data in the big data record and there is no loss to follow-up. The research objects that do not have an outcome at the end of the cohort or cannot continue to be observed due to reasons such as death from non-research outcomes during the process are defined as censored.

[0080] Since in the real-world clinical practice under the background of big data, the complete information of patients is usually recorded during the diagnosis and treatment period rather than on the day when the treatment starts or before the treatment, this leads to a lag in the recording of some covariates, and there is a certain time difference from the index date. According to the OMOP-CDM specification, a covariate capture time window relative to the index date is set in the analysis parameters (such as Figure 4Within 30 days after the indicated date (which can also be 0 or a certain moment before the indicated date), all events before this date can be included in the baseline covariate set, so as to ensure that other concomitant medications and disease diagnoses that may affect the allocation of treatment measures in the initial treatment plan can be fully adjusted.

[0081] The new user cohort under the common data model is usually created based on rules, and the main rules consist of some criteria, including but not limited to: ① domain, that is, the source of storing data in the medical science data warehouse, such as drug domain, condition domain (including disease diagnosis), measurement domain (including laboratory tests), operation domain (surgery, etc.); ② concept set, that is, composed of a series of concepts of clinical characteristics, and different concepts can come from the same or different domains; ③ temporal logic, that is, the time relationship between the included concepts and the occurrence of events, such as it can be set that concepts occurring within one year before a certain exposure event can be included.

[0082] Based on the above criteria, cohorts are created according to specific rule combinations. The main cohort creation rules include: ① enrollment event: The determination of the enrollment event is to determine the time when each individual enters the cohort through a specific event, which is also called the "indicated date". This specific event is usually the exposure factor or treatment measure to be studied, such as drug treatment, disease diagnosis, lifestyle, etc. ② inclusion criteria: Consistent with traditional cohort studies, the specific population of the study is determined by combining the research purpose and the above criteria. ③ exclusion criteria: That is, the rules for withdrawing from the cohort, which can be the occurrence of disease outcomes, the end of the observation period, the end of the research project, the death of the research subject, etc.

[0083] Step S2: Quasi-randomization: Based on 140,000 covariates, construct a high-throughput L1-regularized propensity score matching model to achieve quasi-randomization.

[0084] Specifically include:

[0085] Step 21: Cohort configuration, that is, fuse the target cohort, control cohort and outcome cohort in the constructed new user cohort. For the disease to be analyzed, configure the new user cohort according to the analysis needs of the group or individual and the above cohort creation criteria and rules to generate a cohort for a specific research hypothesis. That is, as described in the previous cohort design, based on the specifications of the common data model, create cohorts for the target exposure cohort (X1), control exposure cohort (X0) and outcome cohort (Y) respectively according to certain rules. The advantage of independent creation is that each created cohort can be repeatedly combined and used, and for different outcomes, the exposure cohort and outcome cohort can be intersected in the analysis program and software for analysis.

[0086] Step 22: Covariate set configuration. Generally, all covariates other than the outcome and exposure are set as the covariate set to be quasi-matched or adjusted; alternatively, the researcher can also reasonably select a subset of covariates based on clinical knowledge or existing evidence for matching or adjustment.

[0087] Step 23: Calculate the high-throughput L1-regularized propensity score to achieve a quasi-randomized design.

[0088] Theoretically, the propensity score is a balancing score. Given the propensity score, the measured baseline covariates have a similar distribution between treated and untreated subjects. The uses of the propensity score include matching subjects with similar PS in the treatment group and the control group, adjusting as a comprehensive covariate, adding weights to subjects through inverse probability of treatment weighting (IPTW), etc. The purpose is to achieve quasi-randomization by controlling confounding bias. Generally speaking, the practical application of PS matching is more extensive, so this design incorporates this method as the main analysis means.

[0089] In RWS, the true propensity score is usually unknown and can be estimated by fitting a L1-regularized regression model of the status of the exposure cohort and the baseline characteristics of the configured 140,000 covariate sets. The exposure prediction probability derived from the regression model is the propensity score, as Figure 5 shown.

[0090] Among them, the calculation formula for L1-regularized propensity score matching is:

[0091] p(T=1|X;θ)=[1 / (1+e -θX )]

[0092]

[0093] where p(T=1|X;θ) is the L1-regularized propensity score, l(θ) is the log-likelihood function, L(θ) is the likelihood function, θ is the model parameter vector, X is the high-dimensional covariate vector, and T i is the binary treatment / treatment indicator variable, and T i is the variable characteristic (0,1) of the i-th individual.

[0094] Step 24: Quasi-randomization diagnosis. Although the regularization propensity score can be completed for matching or weighting between comparison groups through a large number of covariates in the big data cohort warehouse, whether this process can ensure the quality of quasi-randomization needs to be diagnosed through appropriate methods and standards. In this embodiment, two methods are selected and integrated for diagnosis, including: overlap-preference score evaluation and balance diagnosis. When the quasi-randomization is verified, it means that the effective quasi-randomization processing of the new user cohort data is completed.

[0095] (1) Overlap - propensity score evaluation. The basic requirement of the propensity score method is the existence of comparable subjects. Therefore, a key check is to examine the distribution of propensity scores in the two groups. Usually, a "propensity score" propensity score conversion graph can be used to diagnose the overlapping distribution of propensity scores. By plotting the propensity score distribution between the treatment intervention group and the control group, the overlap between the two groups can be observed to guide subsequent propensity score matching or adjustment.

[0096] Taking the comparison of the drug treatment regimens of "metformin" and "other hypoglycemic drugs" for diabetes as an example, the diagnostic process and results of the propensity score are as Figure 6 shown. From the distribution of propensity scores, it can be seen that for many individuals, the treatment they received can be predicted by covariates (discrimination ability AUC = 0.82). However, there are also many overlaps in the distribution of propensity scores, indicating that matching or adjustment can be used to select comparable subjects, suggesting that the treatment group and the non - treatment group are comparable to a certain extent.

[0097] (2) Balance diagnosis. In real - world research, the propensity score is often evaluated by detecting the identifiability of the propensity score model. Among subjects with the same propensity score, the distribution of baseline covariates will tend to be the same between treatment and non - treatment group subjects. Therefore, to evaluate the identifiability of the propensity score model, a judgment on whether the model is visualizable can be made by detecting whether the distribution of baseline covariates between treatment and non - treatment group subjects is similar. Standardized differences are commonly used to compare the means of continuous variables and binary variables between the treatment group and the non - treatment group (multicategorical variables can be represented by a set of binary indicator variables). For a continuous variable, the standardized difference is defined as:

[0098]

[0099] where and represent the sample means of the treatment group and the non - treatment group respectively, while and represent their sample variances respectively;

[0100] For a binary variable, the standardized difference is defined as:

[0101]

[0102] where p 处理 and p 对照 represent the rates of subjects in the treatment group and the non - treatment group respectively;

[0103] Use the distribution map of the difference in standard means before and after propensity score matching to observe the magnitude of the standardized differences of each covariate before and after matching. If the standardized differences of the vast majority of covariates are less than 0.1 after matching, it can be considered that the matching balance is successful. For example, Figure 6 For each covariate (a total of 3,500 covariates) in Figure 6 , the standardized differences before matching are very large (horizontal axis), while the standardized differences after matching are very small (vertical axis), and the absolute value differences after matching are all <0.1, indicating that the effect after matching is good, and it can be considered that the process of quasi-randomization is successful.

[0104] Step S3: Detect the remaining confounding variables and estimation errors through negative control / synthetic positive control.

[0105] Although the bias caused by large-scale observable confounding variables has been controlled after the above-mentioned regularized propensity score matching and diagnosis. However, there may still be some unobserved confounding variables, resulting in unknown bias in the causal effect estimation. In addition, systematic estimation errors such as selection bias and measurement error cannot be controlled by PS either. Therefore, in this embodiment, a negative control and synthetic positive control design strategy is adopted in the causal effect estimation diagnosis to diagnose whether there are unobserved confounding biases and systematic errors.

[0106] Specifically, it includes:

[0107] Step S31: Negative control design. The purpose is to reproduce a condition that does not involve the hypothesized causal mechanism but is likely to involve the same sources of bias that may exist in the exposure-outcome association, mainly including two categories: negative control exposure and negative control outcome, both of which are "exposure-outcome" pairs known to have no causal effect. Among them, let N be the ideal negative control outcome, X and Y be the treatment factor and target outcome to be studied respectively, U be the unobserved confounding variable, and L be the observable confounding variable. Ideally, except that X does not cause N, N should have the same incoming arrows as Y. Under the condition of meeting this standard, N is called "U-comparable" to Y. If N and Y are U-comparable and N is not caused by X, then the negative result obtained from the study of X-N means that the association of X-Y is unlikely to be biased by the path of the negative control passed through.

[0108] In real-world studies, a large number of negative control outcomes should be included as much as possible to incorporate different types of biases to the greatest extent, so as to approach the assumption condition of "U-comparability" between the negative control outcome and the target outcome. The theoretical effect estimate of the target exposure and the negative control outcome is 0, that is, the Hazard Ratio (HR) is 1. Also, in real-world studies, variables that cannot have a causal relationship with the outcome should be selected as negative control exposures, or variables that have no causal relationship with the exposure should be selected as negative control outcomes. Then, after adjusting for known observable confounders, if there is still a causal effect between the exposure and the outcome, it indicates that there are still residual confounding variables.

[0109] Step S32: Synthesize positive control design. Contrary to the negative control, the positive control is an "exposure-outcome" pair with a known causal effect and its effect estimate. By comparing the detected true relative risk (RR less than or greater than 1) with the actual estimate, the validity of the obtained results is evaluated. However, in the real world, there are certain limitations in the selection of positive controls. First, compared with negative controls, positive controls are more difficult to find; second, there are currently no "exposure-outcome" pairs with known causal effect sizes.

[0110] To solve the above problems, in this embodiment, the negative control outcome is transformed into a positive control outcome by adding an additional number of simulated outcomes. The basic principle is that assuming that under a certain treatment measure (X = 1), a total of N negative control outcomes (RR = 1) are observed. If the number of outcomes under this treatment measure is doubled (2N), then theoretically the corresponding risk will double (RR = 2), thus transforming the negative control outcome into a positive control outcome with a known effect size. Using the high-dimensional covariate information in the database, according to the research recommendation, 10-fold cross-validation L1-regularized Poisson regression is used to construct a prediction model for a certain negative control outcome in the treatment group using all available covariate information in the scientific data warehouse.

[0111] Then, based on the prediction model, the predicted probability of each individual in the treatment group is obtained. According to the manually selected "true effect value (RR)", the number of outcomes to be increased is determined, and the corresponding number of samples is changed into the synthesized positive control outcome according to the individual predicted probability. The "positive control outcome" obtained in this way includes the real negative control outcome and the simulated outcome, and maintains the same bias source as the original negative control outcome and the large-scale covariates. The schematic diagram of synthesizing the positive control from the negative control is as Figure 7 shown.

[0112] For negative control variables that are unlikely to have a causal relationship with the outcome or exposure, the outcome event rate is calculated after artificially adding the causal effect through statistical simulation. Then, the difference between the true outcome event rate and the simulated outcome event rate is compared. If the difference is statistically significant, there may still be residual confounding or the model is incorrect.

[0113] Step S4: Calibrate residual confounding and other systematic errors through P value calibration and confidence interval calibration.

[0114] In addition to being used to evaluate the bias in the effect estimates after the propensity score, negative and positive controls also require further empirical calibration of the P value, point estimate, and confidence interval (CI) estimated by the outcome model when possible systematic errors or biases are detected. The specific methods are divided into P value calibration and confidence interval calibration.

[0115] Specifically:

[0116] (1) P value calibration: Most statistical significance tests in medical research are estimated using theoretical null distributions, which require many assumptions to ensure their validity, and these assumptions are often difficult to implement in real-world research. The actual P value only reflects the sampling error caused by the limited sample size, and cannot reflect systematic errors such as confounding bias. Therefore, in the presence of systematic errors, when the test level α = 0.05, the type I error is usually greater than 5%. In real-world research, since systematic errors such as selection bias are inevitable, conclusions drawn based on directly obtained P values ​​often lead to higher type I errors.

[0117] Therefore, to obtain P values ​​that are close to unbiased estimates, empirical null distributions are derived from the actual effect estimates of the negative control, and these distributions are used to recalculate calibrated P values ​​that take into account both random and systematic errors.

[0118] (2) Confidence interval calibration: Similar to P-value calibration, when the null hypothesis H0 is true, in the case of biased estimation, the probability that the true effect value falls within the 95% confidence interval will be less than 95%. At this time, a similar method to P-value calibration can be used to implement confidence interval calibration based on negative control outcomes and synthesized positive control outcomes. The basic principle is to fit a "systematic error model" using the effect estimates of positive and negative controls, and then use this model to calculate the calibrated point estimates and confidence intervals corresponding to the actual effect size. This model assumes that the systematic error follows a Gaussian probability distribution around the true effect size.

[0119] Step S5: In summary, after regularization of PS, P value and confidence interval calibration, the research results can be regarded as causal estimates after controlling known confounding and unknown bias as much as possible, and can be recommended as evidence-based medicine.

[0120] Then, according to the analysis requirements of the population, by selecting an outcome model and based on the dataset of the cohort to be analyzed with quasi-randomization, the estimated value of the average causal effect (ACE) of the population is estimated.

[0121] An outcome model refers to a statistical analysis model used to estimate exposure and outcome. The samples after propensity score processing will be directly used for effect estimation.

[0122] The calculation formula for the estimated value of the average causal effect of the population is:

[0123]

[0124] Among them, n1 is the sample size of the binary treatment variable with a characteristic of 1 (intervention group), n0 is the sample size of the binary treatment variable with a characteristic of 0 (control group), Y is the outcome value of an individual, is the average causal effect of the intervention group, is the average causal effect of the control group.

[0125] In this embodiment, the classical Cox proportional hazards regression model is used as the outcome model. After selecting the outcome model, it is necessary to test the assumption conditions of the model to ensure the correct use of the model. For example, the Cox model needs to conduct a hypothesis test on the proportional hazards assumption of the exposure factor on the outcome, and draw a KM curve, etc., to test whether the curves of the target group and the control group cross or overlap. After the assumptions are met, the estimated value of the average causal effect of the population can be obtained through the outcome model.

[0126] Among them, (1) Propensity score matching. Propensity score matching is to match treated subjects and non-treated subjects with similar PSs so that the confounding variables of the two groups are in balance, thereby obtaining an unbiased estimate of the average treatment effect (ATE) of the treatment / intervention. Commonly used is 1:1 matching, and 1:k matching can also be used. Once the matching is successful, the treatment effect can be estimated by directly comparing the outcome differences between the treated group and the non-treated group. If the outcome is a numerical variable, the treatment effect can be estimated by the difference in the average outcome between the treated group and the non-treated group; if the outcome is a binary variable (for example, depression diagnosis), the treatment effect can be estimated by the difference in the rates between the two groups.

[0127] (2) Inverse probability weighting. First, the information of multiple major confounding variables is integrated into a single variable propensity score; then, the propensity score is used as the confounding variable to be balanced and weighted through the standardization principle to make the propensity score distributions consistent in each comparison group, so as to achieve the purpose of making the distributions of each confounding variable consistent in each comparison group. This method regards each observation unit as a stratum, and different propensity score values indicate different probabilities of this observation unit in the two groups. Under the assumption that there are no unobserved confounding variables, the weighted adjustment is based on the comparison of two opposite events under certain conditions to adjust the data, that is, the two opposite situations of assuming that each subject receives the treatment factor and that each subject does not receive the treatment factor.

[0128] Step S6: According to the individual's analysis requirements, by selecting an outcome model, and based on the dataset of the cohort to be analyzed for quasi-randomization, individual causal effect estimation is performed.

[0129] Among them, counterfactual potential outcome learning is used to estimate the personalized causal effect of the intervention measure.

[0130] The key points of the counterfactual model design are: Based on the idea of mimicking randomization (such as propensity score matching), under the conditional ignorability assumption, causal learning is used to estimate the individualized causal effect (ICE). Let X = 1 represent that the individual is in the exposure group, and X = 0 represent that the individual is in the control group; Y1 and Y0 respectively represent the outcome values observed if the individual is exposed and not exposed to the risk factor / treatment. Then, theoretically, in the counterfactual framework, the individual i's causal effect ICE i = Y1 - Y0; however, since it is impossible for the same individual to observe the values of Y1 and Y0 simultaneously, ICE is usually difficult to obtain.

[0131] Therefore, in this embodiment, causal machine learning algorithms such as the Uplift model (Uplift Modelling, UM) and Bayesian Additive Regression Trees (BART) are used to achieve the estimation of the personalized causal effect of the intervention measure. Its design principle is: Using the above deep learning algorithms, individuals with similar features are learned from a large sample to predict / fill in the potential outcome of the target individual i so as to achieve the personalized causal effect estimation.

[0132] Among them, the process of estimating the individual causal effect using the Uplift model includes:

[0133] Step S61: The Uplift model passes through a deep causal neural network, (such as Figure 8As shown in (a) below, machine learning algorithms such as causal random forests and causal conditional inference forests use machine learning algorithms to achieve counterfactual filling on the premise of a given individual feature marker set X, and predict the values of Y for the sample in the treatment group and the control group, such as Figure 8 in (b) below, and then fit the difference in benefits between the exposed / treatment group and the control group at the individual level to estimate the individual net effect value;

[0134] Step S611: Cohort configuration. Integrate the constructed target cohort, control cohort, and outcome cohort to generate a cohort for a specific research hypothesis.

[0135] Step S612: Covariate set configuration. Use the L1-regularized propensity score model and select covariates through the L1-regularized penalty method; the researcher can also reasonably select a subset of covariates based on clinical knowledge or existing evidence for matching or adjustment.

[0136] Step S613: Model selection, construction, and estimation. Divide the data into a training set and a test set, and construct models MT and MC; furthermore, use the two models respectively, through machine learning algorithms such as deep neural networks, causal random forests, and causal conditional inference forests, to predict the values of Y for the sample in the treatment group and the control group ( Figure 8 in (b) below), and the difference between the two is used as its lift value, which is the individualized causal effect (ICE).

[0137] Step S614: Model evaluation. Use an uplift bar chart, such as Figure 8 in (c) below and an ROC curve such as Figure 8 in (d) below to evaluate the model performance. The uplift model is usually indirectly evaluated by dividing the deciles to align the data of the experimental group and the control group, rather than directly evaluating on a test set. On the test set, the uplifts of the experimental group and the control group are sorted from low to high and divided into ten equal parts, namely deciles, which are top 10%, top 20%,......, top 100%. For each decile in the experimental group and the control group, calculate the mean of E[Y T |X T and E[Y C |X C , that is, the mean of the predicted scores, and then subtract them to obtain the uplift within this decile.

[0138] Step S615: Visualization of individualized efficacy. Such as Figure 8As shown in (e) therein, the PT-PC difference of each sample in the treatment group and the control group is predicted one by one, and a bar chart is drawn; when the difference is greater than 0, the treatment benefits the individual; when the difference is less than 0, the treatment is harmful to the individual; if the difference is 0, the treatment is ineffective for the individual. According to the PT-PC difference of each sample, a decision tree is used for subgroup identification (such as Figure 8 as shown in (f) therein), with the aim of discovering subgroups of people who are suitable or unsuitable for treatment.

[0139] Among them, the individual causal effect estimation process using the BART model includes:

[0140] Step S62: The Bayesian Additive Regression Trees (BART) method is a Bayesian tree method that combines Bayesian theory and the additive tree model. By integrating multiple trees, the model has strong generalization ability.

[0141] Step S621: Cohort configuration. The constructed target cohort, control cohort, and outcome cohort are integrated to generate a cohort for a specific research hypothesis.

[0142] Step S622: Covariate set configuration. Using the L1-regularized propensity score model, covariates are screened through the L1-regularized penalty method to obtain a regularized covariate set; based on this covariate set, the researcher can also further reasonably select a subset of covariates according to clinical knowledge or existing evidence-based evidence for matching or adjustment.

[0143] Step S623: Model selection, construction, and estimation. The data is divided into a training set and a test set. Prior information is added to the important parameters that determine the BART model using the Bayesian method, the joint posterior distribution of the model parameters for the given sample data set is calculated, and parameter values are obtained based on the Bayesian Backfitting MCMC (Markov Chain Monte Carlo) algorithm.

[0144] Step S624: Model evaluation. Bar charts and AUC curves are used to evaluate the model performance.

[0145] Step S625: Visualization of personalized treatment efficacy. A decision tree is used for subgroup identification, with the aim of discovering subgroups of people who are suitable or unsuitable for treatment.

[0146] It should be noted that the acquisition of all data is based on compliance with laws and regulations and user consent, and the data is legally used.

[0147] Example 2

[0148] This example provides a causal effect estimation system based on high-throughput medical science data variables, including:

[0149] A queue construction module, configured to design a new user queue based on high-throughput medical science data, configure the queue for the disease to be analyzed according to the analysis requirements and configuration rules of a group or an individual, and the new user queue includes exposure queues and outcome queues of a target group and a control group;

[0150] A quasi-randomization module, configured to configure a high-throughput covariate set, fit an L1-regularized regression model of the configured exposure queue and the high-throughput covariate set to estimate a propensity score and perform quasi-randomization processing on the configured new user queue, and when the quasi-randomization is verified, obtain a quasi-randomized dataset of the queue to be analyzed;

[0151] A verification module, configured to perform negative control verification on the quasi-randomized dataset of the queue to be analyzed by using exposure-outcome pairs known to have no causal effect, and perform synthetic positive control verification by using exposure-outcome pairs known to have causal effects and effect estimates, so as to verify whether there are unobserved residual confounding variables and estimation errors;

[0152] A causal effect estimation module, configured to perform P-value calibration and confidence interval calibration on the selected outcome prediction model when there are residual confounding variables and estimation errors, and after the calibration is completed, obtain an average causal effect estimate of the group or an individual causal effect estimate according to the quasi-randomized dataset of the queue to be analyzed by using the outcome prediction model.

[0153] It should be noted here that the above modules correspond to the steps described in Embodiment 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in the above Embodiment 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0154] In more embodiments, there is also provided:

[0155] An electronic device, including a memory and a processor as well as computer instructions stored on the memory and running on the processor, and when the computer instructions are run by the processor, the method described in Embodiment 1 is completed. For the sake of brevity, it will not be elaborated here.

[0156] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, field-programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0157] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A part of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0158] A computer-readable storage medium for storing computer instructions, which when executed by a processor, implement the method described in Embodiment 1.

[0159] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0160] A computer program product, including a computer program, which when executed by a processor, implements the method described in Embodiment 1.

[0161] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which are executed in a device on a target real or virtual processor to perform the process / method described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or split as needed. The machine-executable instructions for program modules can be executed locally or within a distributed device. In a distributed device, program modules can be located in local and remote storage media.

[0162] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, such that when the program code is executed by the computer or other programmable data processing devices, the functions / operations specified in the flowchart and / or block diagram are implemented. The program code can be executed entirely on the computer, partially on the computer, as an independent software package, partially on the computer and partially on a remote computer, or entirely on a remote computer or server.

[0163] In the context of the present invention, the computer program code or relevant data can be carried by any suitable carrier so that a device, apparatus or processor can perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals can include electrical, optical, radio, acoustic or other forms of propagated signals, such as carrier waves, infrared signals, etc.

[0164] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with this embodiment can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of this application.

[0165] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, this is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A causal effect estimation method based on high-throughput medical science data variables, characterized in that: include: Designing a new user cohort based on high-throughput medical science data, and configuring the cohort according to the analysis requirements and configuration rules of the group or individual for the disease to be analyzed, wherein the new user cohort includes an exposure cohort and an outcome cohort of a target group and a control group; A high-throughput covariate set is configured, and an L1 regularized regression model of the configured exposure cohort and the high-throughput covariate set is fitted to estimate the propensity score and quasi-randomize the configured new user cohort. After passing the quasi-randomization verification, a quasi-randomized cohort data set to be analyzed is obtained; For the randomized cohort data set to be analyzed, negative control validation was performed using exposure-outcome pairs that were known to have no causal effect, and synthetic positive control validation was performed using exposure-outcome pairs that were known to have a causal effect and effect estimate, to verify whether there were unobserved residual confounding variables and estimation errors; When there are residual confounding variables and estimation errors, the selected outcome prediction model is calibrated for P value and confidence interval. After calibration, the outcome prediction model is used according to the quasi-randomized cohort data set to be analyzed to obtain the average causal effect estimate of the group or the individual causal effect estimate.

2. The causal effect estimation method based on high-throughput medical science data variables according to claim 1, characterized in that: The process of designing a new user cohort and configuring the cohort includes: establishing a medical science data warehouse and defining a concept set based on high-throughput medical science data, extracting data for cohort configuration according to the set standards and configuration rules; Among them, the standards include: domain, which is the source of data stored in the medical science data warehouse; concept set, which is the concept composition of clinical characteristics, and different concepts come from the same or different domains; temporal logic, which is the temporal relationship between the included concepts and the occurrence of events; configuration rules include enrollment events, inclusion criteria and exclusion criteria; The process of obtaining the quasi-randomized cohort data set to be analyzed includes: setting all covariates except outcomes and exposures as covariate sets, and using overlap-preference score evaluation and balance diagnosis to perform quasi-randomization verification.

3. The causal effect estimation method based on high-throughput medical science data variables according to claim 1, characterized in that: Negative control validation includes: selecting a variable that has no causal effect relationship with the outcome as a negative control exposure, or selecting a variable that has no causal effect relationship with the exposure as a negative control outcome. After adjusting the known observable confounding variables, if there is still a causal effect relationship between the exposure and the outcome, it means that there are still residual confounding variables.

4. The causal effect estimation method based on high-throughput medical science data variables according to claim 1, characterized in that: Synthetic positive control validation includes: By adding an additional number of simulated outcomes to the negative control outcomes, the negative control outcomes are transformed into positive control outcomes; wherein, a prediction model for the negative control outcomes is constructed using a high-throughput covariate set, and the individual prediction probability is obtained based on the prediction model. According to the set true effect value, the number of outcomes to be added is determined, thereby converting the corresponding number of negative control outcomes into synthetic positive control outcomes according to the individual prediction probability. The synthetic positive control outcomes include the real negative control outcomes and the simulated outcomes, and are consistent with the bias source between the original negative control outcomes and the covariates; For negative control variables that have no causal effect relationship with the outcome or exposure, the simulated outcome event rate is calculated after artificially adding the causal effect; the difference between the true outcome event rate and the simulated outcome event rate is compared to determine whether there are residual confounding factors or estimation errors.

5. The causal effect estimation method based on high-throughput medical science data variables according to claim 1, characterized in that: P value calibration includes: deriving an empirical null distribution from the actual effect estimate of the negative control, and recalculating the calibrated P value to obtain the P value under unbiased estimation; Confidence interval calibration includes: achieving confidence interval calibration based on negative control outcomes and synthetic positive control outcomes; specifically, fitting the systematic error model through the effect estimates of the positive and negative controls, and using the systematic error model to calculate the calibrated point estimate and confidence interval corresponding to the actual effect estimate.

6. The causal effect estimation method based on high-throughput medical science data variables according to claim 1, characterized in that: The process of estimating the average causal effect estimate of the group includes: using the Cox proportional hazard regression model as the selected outcome prediction model, and obtaining the average causal effect estimate of the group after testing the assumptions of the outcome prediction model; The process of estimating the individual causal effect estimate includes: estimating the individual causal effect estimate based on counterfactual potential outcome learning.

7. A causal effect estimation system based on high-throughput medical science data variables, characterized in that: include: A cohort construction module is configured to design a new user cohort based on high-throughput medical science data, and configure the cohort according to the analysis requirements and configuration rules of a group or individual for the disease to be analyzed, wherein the new user cohort includes an exposure cohort and an outcome cohort of a target group and a control group; The quasi-randomization module is configured to configure a high-throughput covariate set, fit an L1 regularized regression model of the configured exposure cohort and the high-throughput covariate set, estimate the propensity score, and quasi-randomize the configured new user cohort. After passing the quasi-randomization verification, a quasi-randomized cohort data set to be analyzed is obtained; A validation module is configured to perform negative control validation on the quasi-randomized cohort data set to be analyzed using exposure-outcome pairs that are known to have no causal effect, and to perform synthetic positive control validation on exposure-outcome pairs that are known to have a causal effect and an effect estimate, so as to verify whether there are unobserved residual confounding variables and estimation errors; The causal effect estimation module is configured to perform P value calibration and confidence interval calibration on the selected outcome prediction model when there are residual confounding variables and estimation errors. After the calibration is completed, the average causal effect estimate of the group or the individual causal effect estimate is obtained by using the outcome prediction model based on the quasi-randomized cohort data set to be analyzed.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images