A big data-based chronic disease quality control data analysis processing system and method

Abstract

This invention relates to the field of medical big data analysis technology, and discloses a system and method for analyzing and processing chronic disease quality control data based on big data. The method includes: acquiring a panel dataset of staged quality control indicators and a table of missing data patterns; classifying missing events into systematic, capability, and managerial missing events using an association rule mining algorithm; performing differentiated processing on different categories of missing data and recording the effective sample size; generating a directed acyclic graph of causality before and after intervention with confidence level labels; calculating the probability of false positives due to missing data and generating a list of causal structure differences; generating a counterfactual baseline sequence; discounting and correcting the causal transmission effect components, decomposing the difference sequence into four components; and generating a comprehensive analysis report.

Description

Technical Field

[0001] This invention relates to the field of medical big data analysis technology, specifically to a system and method for analyzing and processing chronic disease quality control data based on big data. Background Technology

[0002] In chronic disease quality control management, management departments collect panel data of quality control indicators reported by multiple medical institutions over multiple assessment cycles. After implementing intervention measures, they conduct causal analysis on large-scale quality control data before and after the intervention to evaluate the intervention effect. Existing methods typically generate causal directed acyclic graphs based on panel data of quality control indicators before and after the intervention. By comparing the differences in causal structure before and after the intervention, they identify the impact of the intervention on the causal relationship between indicators and decompose the improvement of indicators into direct intervention effects, indirect effects of causal transmission, and contributions from external factors for attribution analysis.

[0003] However, in the actual collection of large-scale chronic disease quality control data, the data completeness before and after the intervention often exhibits a significant asymmetry. In the early stages of intervention, due to incomplete historical information systems and varying data management capabilities among institutions, the reported data generally contains many gaps. In the later stages of intervention, data is more complete because intervention measures are usually accompanied by strengthened data management. This asymmetry in data completeness leads to worse conditions for generating causal directed acyclic graphs (DAGs) in the early stages of intervention compared to the later stages: before intervention, some indicator pairs fail to detect true causal edges due to insufficient effective sample size, while in the later stages, the same indicator pairs are identified as causal edges due to sufficient data. Existing methods, when comparing structural differences, misjudge these detected differences caused by data quality variations as newly added causal edges, incorrectly attributing them to the intervention altering the causal structure. This leads to an overestimation of the causal transmission effect, severely distorting the attribution conclusions of the intervention effect. Summary of the Invention

[0004] This invention provides a data analysis and processing system and method for chronic disease quality control based on big data, which solves the technical problems in related technologies such as the inability to distinguish the causes of missing data, the difficulty in identifying the causal structure due to missing data, and the inaccuracy of intervention effect estimation due to interference from data quality when performing attribution analysis of the intervention effect in chronic disease quality control.

[0005] This invention discloses a method for analyzing and processing chronic disease quality control data based on big data, comprising: Acquire panel data of quality control indicators from multiple medical institutions before and after the intervention, divide them into pre-intervention datasets and post-intervention datasets according to the intervention time nodes, and statistically analyze the effective data coverage and missing pattern characteristics of each indicator in each period to generate a staged quality control indicator panel dataset and a staged data missing pattern characteristic table. For the phased data missing pattern feature table, the association rule mining algorithm is used to analyze the relationship between missing events and institutional attributes and time factors, and each missing event is classified into three cause categories: systemic missing, capability missing and management missing, and the missing cause classification labeling results are generated. Based on the classification and labeling results of the missing causes, differential missing data processing is performed on the pre-intervention dataset and the post-intervention dataset. Systematic missing data is filled using linear extrapolation, while capability missing data and management missing data are marked as excluded. A staged dataset and an effective sample size record table are generated after missing data processing. Causal tests are applied to each stage dataset after missing data processing to generate causal directed acyclic graphs before and after intervention. Structural differences are compared between the causal directed acyclic graphs before and after intervention. Based on the effective sample size record table, the probability of missing data causing false positives is calculated for each difference edge, and a list of causal structural differences with missing data causing false positives is generated. Based on the causal structure of the pre-intervention causal directed acyclic graph and the time trend of the pre-intervention data, counterfactual extrapolation is performed on the target indicators to generate a counterfactual baseline sequence. The difference sequence between the actual value sequence after intervention of the target indicator and the counterfactual baseline sequence is calculated. Causal edges involving missing or falsified risks on the causal path are corrected by discount weights. The difference sequence is decomposed into direct intervention effect component, corrected causal transmission effect component, data quality interference component, and external factor residual component to generate a four-component corrected attribution result.

[0006] Furthermore, the method of using association rule mining algorithms to analyze the relationship between missing events and organizational attributes and time factors, classifying each missing event into three categories: systemic missing, capability missing, and management missing, includes: extracting missing event records from the phased data missing pattern feature table, each missing event record containing a missing indicator identifier, a missing organization identifier, a missing period identifier, and corresponding organizational attribute features, including organization level, region, and information system version; using missing event records as transaction sets and organizational attribute features and time factors as item sets, applying association rule mining algorithms to extract a set of association rules that meet preset minimum support and minimum confidence; when the association rule corresponding to a missing event points to a batch of missing information system versions within a specific time period, the missing event is marked as a systemic missing; when the association rule corresponding to a missing event points to a persistent missing indicator for an organization of a specific organization level, the missing event is marked as a capability missing; when a missing event does not meet the matching conditions of the above two types of association rules, the missing event is marked as a management missing.

[0007] Furthermore, the method of filling systematic missing data using linear extrapolation includes: for missing events marked as systematic missing, obtaining the sequence of observations for the most recent available period for the corresponding institution on the corresponding indicator, calculating the filling value for the missing period using linear extrapolation, and writing the filling value into the missing position in the corresponding dataset; for missing events marked as capability-related or management-related missing, no filling is performed, and all indicator data of the corresponding institution in that period are marked as excluded; the effective sample size record table records the number of effective observations for each pair of indicator combinations in the early and late stages of the intervention.

[0008] Furthermore, the step of applying causal tests to each stage dataset after missing data processing to generate causal directed acyclic graphs before and after intervention includes: applying Granger causality tests based on vector autoregression models to generate a set of candidate causal edges for each stage dataset after missing data processing; then applying the PC algorithm to generate a causal directed acyclic graph after conditional independence tests; recording the p-value of the statistical test, the causal effect strength coefficient, and the corresponding effective sample size for each causal edge, where the causal effect strength coefficient is the regression coefficient of the corresponding lag term in the vector autoregression model; and attaching the above information to each edge of the causal directed acyclic graph to generate causal directed acyclic graphs before and after intervention with test confidence labels.

[0009] Further, the calculation of the missing sample probability for each dissimilar edge based on the effective sample size record table includes: marking edges that exist only in the post-intervention causal directed acyclic graph as newly added causal edges, marking edges that exist only in the pre-intervention causal directed acyclic graph as disappeared causal edges, and marking edges that exist in both graphs but whose causal effect strength coefficient changes by more than a preset change threshold as effect strength change edges; for newly added causal edges, obtaining the effective sample size of the corresponding indicator in the pre-intervention period and the minimum sample size threshold required for causal testing; when the effective sample size in the pre-intervention period is less than the minimum sample size threshold, the missing sample probability is the minimum sample size threshold. The ratio of the difference between the threshold value and the effective sample size in the early intervention period to the minimum sample size threshold value; when the effective sample size in the early intervention period is not less than the minimum sample size threshold value, the probability of missing data is zero; for disappearing causal edges, obtain the effective sample size of the corresponding indicator pair in the later intervention period, and calculate the probability of missing data in the same way as adding causal edges; for edge changes in effect intensity, when the effective sample size in either the early or late intervention period is lower than the minimum sample size threshold value, calculate the probability of missing data in the above way; when the effective sample size in both periods is not lower than the minimum sample size threshold value, the probability of missing data is zero.

[0010] Furthermore, the step of counterfactually extrapolating the target indicator based on the causal structure of the pre-intervention causal directed acyclic graph and the time trend of the pre-intervention data to generate a counterfactual baseline sequence includes: extracting the set of parent nodes of the target indicator and the corresponding causal effect strength coefficients from the pre-intervention causal directed acyclic graph; calculating the extrapolated expected value of each parent node indicator in each period after the intervention using linear extrapolation; weighting and summing the extrapolated expected value of each parent node indicator with the corresponding causal effect strength coefficient, superimposing the intercept term obtained by fitting the pre-intervention data, calculating the counterfactual expected value of the target indicator in each period after the intervention, and summing the counterfactual expected values ​​of each period to generate a counterfactual baseline sequence; wherein, the time series of each parent node indicator is normalized to the mean based on the range before inputting the linear extrapolation, and the causal effect strength coefficient is estimated based on the normalized data.

[0011] Furthermore, the step of correcting causal edges involving missing or falsified risk on the causal path by applying a discount weight includes: calculating the original component of the causal transmission effect based on the causal path from the directly affected indicator to the target indicator in the post-intervention causal directed acyclic graph; querying the causal structure difference list labeled with missing or falsified risk for the causal edges involved in the causal path; for the causal edges labeled as missing or falsified, multiplying the original causal effect strength coefficient by one and subtracting the discount weight obtained from the corresponding missing or falsified probability to obtain the corrected causal effect strength coefficient; recalculating the causal transmission effect component based on the corrected causal effect strength coefficient to generate the corrected causal transmission effect component; wherein, the data quality interference component is the difference between the original causal transmission effect component and the corrected causal transmission effect component, and the external factor residual component is the remaining difference in the difference sequence after deducting the direct intervention effect component, the corrected causal transmission effect component, and the data quality interference component.

[0012] Furthermore, the calculation method for the direct intervention effect component is as follows: obtain the node set of the direct intervention effect indicators in the causal directed acyclic graph after intervention, and sum the differences between the actual observed values ​​of each indicator in the node set of the direct intervention effect indicators and the corresponding expected values ​​in the counterfactual baseline sequence in each intervention period, according to the direct causal effect intensity coefficient of each indicator on the target indicator.

[0013] Furthermore, it also includes: calculating the confidence interval of each component using Bootstrap resampling on the four-component corrected attribution results; calculating the proportion of the data quality interference component in the total effect, and performing a reliability rating based on this proportion. When the proportion is lower than a first preset threshold, it is rated as highly reliable; when the proportion is between the first preset threshold and a second preset threshold, it is rated as moderately reliable; when the proportion exceeds the second preset threshold, it is rated as low reliable and triggers an attribution reliability insufficient warning; summarizing the indicator pairs corresponding to the difference edges with a probability of missing falsehood greater than zero in the causal structure difference list with missing falsehood risk labeling, the corresponding missing cause categories in the missing cause classification labeling results, and the institutional identifiers involved in management missingness as data quality improvement suggestions; integrating and outputting the corrected attribution estimate, the confidence interval of each component, the reliability rating, and the data quality improvement suggestions to generate a comprehensive analysis report on the effect of chronic disease quality control intervention.

[0014] This invention discloses a big data-based chronic disease quality control data analysis and processing system, used to execute the aforementioned big data-based chronic disease quality control data analysis and processing method, comprising: The phased data acquisition module is used to acquire panel data of quality control indicators from multiple medical institutions before and after the intervention. It is divided into pre-intervention datasets and post-intervention datasets according to the intervention time nodes, and the effective data coverage and missing pattern characteristics of each indicator in each period are statistically analyzed to generate phased quality control indicator panel datasets and phased data missing pattern characteristic tables. The missing cause classification module is used to analyze the relationship between missing events and organizational attributes and time factors in the missing pattern feature table of the phased data using association rule mining algorithm, classify each missing event into three cause categories: systemic missing, capability missing and management missing, and generate missing cause classification labeling results; The differential missing data processing module is used to perform differential missing data processing on the pre-intervention dataset and the post-intervention dataset based on the missing data cause classification and labeling results. Systematic missing data is filled using linear extrapolation, while capability missing data and management missing data are marked as excluded. The module generates a staged dataset and an effective sample size record table after missing data processing. The causal structure difference analysis module is used to apply causal tests to the datasets of each stage after missing data processing to generate causal directed acyclic graphs before and after intervention. The module compares the structural differences between the causal directed acyclic graphs before and after intervention, calculates the probability of missing data causing false positives for each difference edge based on the effective sample size record table, and generates a list of causal structure differences with missing data causing false positives labeling. The counterfactual extrapolation module is used to perform counterfactual extrapolation on the target indicators based on the causal structure of the pre-intervention causal directed acyclic graph and the time trend of the pre-intervention data, and generate a counterfactual baseline sequence. The modified attribution decomposition module is used to calculate the difference sequence between the actual value sequence after intervention of the target indicator and the counterfactual baseline sequence. It modifies the causal edges involving the risk of missing data and falsification on the causal path according to the discount weight, and decomposes the difference sequence into the direct intervention effect component, the modified causal transmission effect component, the data quality interference component, and the external factor residual component, generating a four-component modified attribution result.

[0015] The present invention has the following beneficial effects: This invention classifies missing events into systematic, capability, and managerial missing events through association rule mining, and employs a differentiated processing strategy for each category—linear extrapolation is used to fill in systematic missing events, while capability and managerial missing events are excluded. This differentiated processing ensures that causal tests are performed based on reliable data in both the pre- and post-intervention periods, avoiding situations where simple uniform filling masks data management deficiencies or ignores missing events, leading to inconsistencies in causal test conditions before and after the intervention.

[0016] This invention calculates the probability of missing or falsified edges for each differing edge. This transforms the determination of differences in causal structure before and after intervention from a binary judgment to a probabilistic assessment. In the attribution decomposition stage, discount weights are used... The causal edges involving the risk of missing data are corrected so that the spurious increase in the causal transmission effect component due to data quality differences is separated into an independent data quality interference component. ,in This represents the probability of spurious errors due to missing data corresponding to causal edges. Removing data quality interference components enables attribution conclusions to distinguish between the causal structure changes truly caused by interventions and newly detected existing causal relationships after data improvement, thereby suppressing the interference of spurious structural changes on attribution conclusions under conditions of asymmetric data completeness.

[0017] This invention provides management departments with a quantitative reference for the reliability of attribution conclusions by assessing the proportion of data quality interference components in the total effect and outputting a reliability rating. Furthermore, it generates data quality improvement suggestions by summarizing the indicator pairs and institutional details involved in missing false edges, thus providing a clear direction for subsequent improvements in data collection and management. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for analyzing and processing chronic disease quality control data based on big data, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram comparing the effective data coverage of each quality control indicator before and after intervention, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram comparing the number of cycles of missing quality control indicators before and after intervention, provided in an embodiment of the present invention. Figure 4This is a schematic diagram showing the comparison of the effective sample size and minimum threshold before and after intervention based on the key indicators provided in the embodiments of the present invention; Figure 5 This is a schematic diagram comparing the intensity coefficients of the causal side effect before and after intervention related to the HbA1c target achievement rate, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram comparing the actual observed value and the counterfactual expected value of the HbA1c compliance rate provided in this embodiment of the invention; Figure 7 This is a schematic diagram of the four-component corrected attribution results for each cycle after intervention provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the scatter distribution of the probability of falsehood due to missing newly added causal edges and the effective sample size before intervention, provided in an embodiment of the present invention. Figure 9 This is a schematic diagram showing the periodic change of the proportion of data quality interference components to the total difference provided in the embodiments of the present invention. Detailed Implementation

[0019] Example In the big data analytics scenario of chronic disease quality control management, management departments collect panel data of quality control indicators reported by multiple medical institutions over multiple assessment cycles. After implementing intervention measures, they conduct causal analysis on large-scale quality control data before and after the intervention to evaluate the intervention effect. Existing methods typically generate causal directed acyclic graphs based on panel data of quality control indicators before and after the intervention. By comparing the differences in causal structure before and after the intervention, they identify the impact of the intervention on the causal relationship between indicators and decompose the improvement of indicators into direct intervention effects, indirect effects of causal transmission, and contributions from external factors for attribution analysis.

[0020] However, in the actual collection of large-scale chronic disease quality control data, the data completeness before and after the intervention often exhibits a significant asymmetry. In the early stages of intervention, due to incomplete historical information systems and varying data management capabilities among institutions, the reported data generally contains many gaps. In the later stages of intervention, data is more complete because intervention measures are usually accompanied by strengthened data management. This asymmetry in data completeness leads to worse conditions for generating causal directed acyclic graphs (DAGs) in the early stages of intervention compared to the later stages: before intervention, some indicator pairs fail to detect true causal edges due to insufficient effective sample size, while in the later stages, the same indicator pairs are identified as causal edges due to sufficient data. Structural difference comparisons may misjudge these detected differences caused by data quality variations as newly added causal edges, incorrectly attributing them to the intervention altering the causal structure. This results in an overestimation of the causal transmission effect, severely distorting the attribution conclusions regarding the intervention's effectiveness.

[0021] This embodiment provides a data analysis and processing system and method for chronic disease quality control based on big data. By classifying and differentiating the causes of missing data before and after intervention, it quantifies the risk of spurious changes in causal structure caused by missing data, and corrects the attribution conclusions accordingly, thereby obtaining reliable attribution results of intervention effects under the condition of asymmetric data completeness.

[0022] The method for analyzing and processing chronic disease quality control data based on big data in this embodiment is described in [reference needed]. Figure 1 This includes the following steps: Step 1: Obtain the panel dataset of phased quality control indicators and the feature table of missing data patterns.

[0023] Obtain continuous data from multiple medical institutions before intervention implementation. Each cycle and continuous after intervention implementation Panel data of quality control indicators for each period, including The number of assessment periods is preset. The quality control indicator panel data is divided into pre-intervention and post-intervention datasets based on the intervention time point. For each institution in each dataset, the data integrity status of each indicator in each assessment period is labeled. The effective data coverage and missing data pattern characteristics of each indicator in each period are statistically analyzed, generating a phased quality control indicator panel dataset and a corresponding phased data missing data pattern characteristic table.

[0024] It should be noted that data integrity status includes two labels: data complete and data missing. Effective data coverage refers to the ratio of the number of institutional cycles with valid observations for a certain indicator within a specific period to the total number of institutional cycles within that period. Missing data pattern characteristics include the identifier of the missing indicator, the time range of the missing data, the set of institutions involved in the missing data, and whether the missing data exhibits continuity over time.

[0025] Step 2: Generate the classification and labeling results of the missing reasons.

[0026] For the feature table of missing patterns in phased data, the association rule mining algorithm is used to analyze the relationship between missing events and institutional attributes and time factors. Each missing event is classified into three cause categories, and the missing cause classification labeling results are generated.

[0027] Step 2 may specifically include: Step S201: Extract missing event records from the missing pattern feature table of phased data. Each missing event record includes a missing indicator identifier, a missing organization identifier, a missing period identifier, and the corresponding organization attribute features. The organization attribute features include organization level, region, and information system version. Among them, organization level and information system version are categorized data, which are converted into numerical representations using encoding methods before being input into the association rule mining algorithm.

[0028] Step S202: Using missing event records as the transaction set and organizational attribute features and time factors as itemsets, an association rule mining algorithm is applied to extract association rules that satisfy preset minimum support and minimum confidence. The input of the association rule mining algorithm is the missing event transaction set, and the output is the set of association rules that satisfy the threshold conditions.

[0029] Step S203: Classify the causes of missing events based on the set of association rules. When the association rule corresponding to a missing event points to a batch of missing data in a specific information system version within a specific time period, the missing event is marked as a systemic missing event; when the association rule corresponding to a missing event points to a persistent missing data in a specific indicator for an organization at a specific organizational level, the missing event is marked as a capability missing event; when a missing event does not meet the matching conditions of the above two types of association rules, the missing event is marked as a management missing event. Summarize the cause categories of each missing event to generate the missing cause classification and labeling results.

[0030] It should be noted that systemic data loss refers to the loss of large batches of data due to system-level reasons such as information system upgrades or interface changes. Systemic data loss is characterized by affecting the same set of indicators across multiple organizations using the same system version within a specific time window. Capability-related data loss refers to data loss due to insufficient institutional testing capabilities, resulting in the inability to collect specific indicators. Capability-related data loss is characterized by its continuous impact on specific indicators of specific levels of organizations over a relatively long period. Management-related data loss refers to scattered data loss due to oversights in the reporting process. The timing and institutional distribution of management-related data loss do not exhibit the patterns described above.

[0031] Step 3: Generate the phased dataset after missing data processing and the effective sample size record table.

[0032] Based on the classification and labeling results of missing causes, differential missing data processing was performed on the pre-intervention dataset and the post-intervention dataset, generating the pre-intervention dataset after missing data processing, the post-intervention dataset after missing data processing, and a record table of effective sample size.

[0033] Step 3 may specifically include: Step S301: For missing events labeled as systematically missing, obtain the sequence of observations for the most recent available period on the corresponding indicator for the corresponding institution, calculate the imputation value for the missing period using linear extrapolation, and write the imputation value into the missing location in the corresponding dataset. The input to linear extrapolation is the sequence of most recent available continuous observations, and the output is the extrapolated value for the time point corresponding to the missing period.

[0034] Step S302: For missing events marked as capability-related or management-related deficiencies, no filling process is performed. Instead, all indicator data of the corresponding institution in that period are marked as excluded, so that they are not included in subsequent calculations.

[0035] Step S303: Calculate the effective sample size for each indicator pair in each period of the pre-intervention and post-intervention datasets after differential missing data processing, and generate an effective sample size record table. The effective sample size record table records the number of effective observations for each indicator pair in the pre-intervention and post-intervention periods.

[0036] It should be noted that the rationale for using linear extrapolation to fill systematic gaps is that these gaps arise from technical interruptions at the information system level. The actual observed values ​​of the corresponding indicators by the relevant institutions continue to change along their original trends during the gap period. Therefore, extrapolated values ​​based on existing observation sequences can reasonably approximate the actual values ​​during the gap period. Capability and management gaps are excluded instead of being filled because they reflect a lack of reliable data collection conditions for the corresponding indicators by the relevant institutions. Filled values ​​cannot reasonably approximate the actual values, and including them in the calculation would introduce bias.

[0037] Step 4: Generate a directed acyclic graph of causality before and after the intervention with confidence level labels.

[0038] For the pre-intervention dataset after missing data processing, Granger causality test based on a vector autoregression model and PC algorithm were applied to generate a directed acyclic graph (DAG) of causality for the pre-intervention quality control indicators. The same operation was performed on the post-intervention dataset after missing data processing to generate a DAG of causality for the post-intervention quality control indicators.

[0039] The Granger causality test takes the time series of each indicator in the staged dataset after missing data processing as input, and outputs the causal direction determination between each pair of indicators and the corresponding statistical test. The PC algorithm takes the set of candidate causal edges and the phased dataset as inputs from the Granger causality test, and outputs a causal directed acyclic graph constrained by the conditional independence test.

[0040] During the causality test, the statistical test results for each candidate causal edge are recorded. The values, causal effect strength coefficients, and corresponding effective sample sizes are obtained. The causal effect strength coefficient is the regression coefficient of the corresponding lagged term in the vector autoregression model. This information is appended to each edge of the causal directed acyclic graph (DAG), generating a pre-intervention DAG with confidence level labels and a post-intervention DAG with confidence level labels.

[0041] Step 5: Generate a list of causal structure differences with missing risk labels.

[0042] Structural differences were compared between two causal directed acyclic graphs before and after the intervention. Three types of changes were identified: newly added causal edges, disappeared causal edges, and edges with changes in effect intensity. The probability of missing or falsified edges was calculated for each difference edge, and a list of causal structural differences with missing or falsified risk labels was generated.

[0043] Step 5 may specifically include: Step S501: Compare the edge set of the post-intervention causal directed acyclic graph with the edge set of the pre-intervention causal directed acyclic graph. Mark edges that exist only in the post-intervention causal directed acyclic graph as newly added causal edges, edges that exist only in the pre-intervention causal directed acyclic graph as disappeared causal edges, and edges that exist in both graphs but whose causal effect intensity coefficient changes by more than a preset change threshold as effect intensity change edges, generating a list of causal structure differences.

[0044] Step S502: For each difference edge in the causal structure difference list, calculate the probability of missing data causing spurious results by combining it with the effective sample size record table.

[0045] For newly added causal edges, obtain the effective sample size of the corresponding indicator pairs in the early stage of intervention. and the minimum sample size threshold required for causality tests ,in This indicator represents the number of samples with valid observations in the early stages of the intervention. The significance level is determined based on the causal testing method used and the pre-set significance level. At that time, there is a possibility that newly added causal edges may not be detected due to insufficient sample size in the early stages of intervention, resulting in a lack of spurious probability. Calculate as follows: in, To increase the effective sample size of the newly added causal edge correspondence indicator in the early stage of intervention, This is the minimum sample size threshold required for causality testing. hour, The value is zero.

[0046] For vanishing causal edges, perform a symmetry check to obtain the effective sample size of the corresponding indicator pair in the later stage of the intervention, and calculate the probability of missing data in the same way.

[0047] For the effect strength change edge, when the effective sample size in either the early or late stage of the intervention is less than Similarly, the probability of false positives due to missing data is calculated in the same way as above; when the effective sample size in both the early and late stages of the intervention is not less than hour, The value is zero.

[0048] It should be noted that, The value of is between zero and one. The closer it is to one, the greater the likelihood that the corresponding difference edge is a spurious change due to missing data; A value of zero indicates that the effective sample size of the corresponding difference edge in both periods meets the causal test requirements, and the corresponding difference edge reflects the real causal structure change brought about by the intervention.

[0049] Step S503: Append the probability of missing data to the corresponding difference edge in the causal structure difference list to generate a causal structure difference list with missing data risk labeling.

[0050] Step 6: Generate the counterfactual baseline sequence.

[0051] Based on the causal structure of the pre-intervention causal directed acyclic graph and the time trend of the pre-intervention data, counterfactual extrapolation is performed on the target indicators targeted by the intervention measures. The expected value sequence of the target indicators in each period after the intervention is calculated under the assumption that no intervention was implemented, and a counterfactual baseline sequence is generated.

[0052] Step 6 may specifically include: Step S601: Extract the set of parent nodes of the target indicator and the corresponding causal effect intensity coefficient from the pre-intervention causal directed acyclic graph, and obtain the time series of the target indicator and its parent node indicators in the pre-intervention dataset after missing data processing.

[0053] Step S602: Based on the time series of the parent node indicators of the target indicator in the pre-intervention period, calculate the expected value of each parent node indicator in each period after the intervention using linear extrapolation. The input of linear extrapolation is the time series of each parent node indicator in the pre-intervention period, and the output is the extrapolated expected value of each parent node indicator at the corresponding time point in each period after the intervention.

[0054] Step S603: The extrapolated expected values ​​of each parent node indicator are weighted and summed with the corresponding causal effect strength coefficients in the pre-intervention directed acyclic graph to calculate the expected values ​​of the target indicator in each post-intervention period. The target indicator in the [number]th [period]... Counterfactual expected value after one intervention cycle Calculate as follows: in, Let be the set of parent nodes in the causal directed acyclic graph of the target indicator before intervention. set of parent nodes The parent node index in the data. For the first The causal effect strength coefficient corresponding to each parent node indicator For the first The parent node index is at the 1st Extrapolated expected value for each intervention cycle To intervene in the intercept term obtained by fitting the previous data, The value range is from the first intervention cycle to the [number missing]. One intervention cycle, The preset number of evaluation periods. The counterfactual expected value for each post-intervention period. The data is then aggregated to generate a counterfactual baseline sequence. This counterfactual baseline sequence represents the expected trajectory of the target indicator under conditions where no intervention was implemented and the causal structure remained as it was before the intervention.

[0055] It should be noted that the above weighted summation operation requires... and The dimensions are compatible. Before linear extrapolation of the time series of each parent node index, mean normalization based on the range is performed to uniformly scale the values ​​of each index to the same range, eliminating the influence of the difference in dimensions of different indicators on the weighted summation result; When fitting the vector autoregressive model, the estimation is based on the normalized data, and... Maintain consistency in dimensions.

[0056] Step 7: Generate four-component corrected attribution results.

[0057] The difference sequence between the actual value sequence after intervention of the target indicator and the counterfactual baseline sequence is calculated. The causal transmission effect component is corrected by combining the causal structure difference list with missing spurious risk labeling. The difference sequence is decomposed into four components to generate four-component corrected attribution results.

[0058] Step 7 may specifically include: Step S701: Calculate the difference between the actual observed values ​​of the target indicator in each period after the intervention and the expected values ​​in the corresponding period in the counterfactual baseline sequence, and generate a difference sequence. ,in For the first The difference between each intervention period, The value range is from the first intervention cycle to the [number missing]. One intervention cycle, This represents the preset number of evaluation cycles.

[0059] Step S702: Based on the causal path from the directly affected indicator to the target indicator in the post-intervention causal directed acyclic graph, calculate the original components of the causal transmission effect. For the causal edges involved in the causal path, query the list of causal structure differences marked with missing or falsified risk, and weight the differences marked as missing or falsified according to a discount weight. For the first... The indicator to the first For each indicator's causal edge, if the causal edge is marked as missing or falsifiable in a list of causal structure differences with missing / falsifiable risk labels, the corrected causal effect strength coefficient... Calculate as follows: in, Let this be the original causal effect strength coefficient of the causal edge in the causal directed acyclic graph after intervention. This represents the probability of missing information causing spurious results for the corresponding causal edge. This serves as the starting index for the causal path. This serves as the termination index for the causal path. The causal transmission effect components are recalculated based on the corrected causal effect strength coefficient, generating the corrected causal transmission effect components. .

[0060] Step S703: Decompose the difference sequence into four components. For each cycle after intervention... The decomposition relationship is as follows: in, For the first The difference between each intervention period, The direct intervention effect component represents the direct effect of the intervention on the target indicator; The component representing the causal transmission effect after correction is denoted as , which indicates the indirect transmission effect of the causal path after discount weighting correction. The data quality interference component represents the difference in effect caused by the false edge discount due to missing data, i.e., the difference between the original causal transmission effect component and the corrected causal transmission effect component. The external factor residual component represents the remaining difference after deducting the above three components; The value range is from the first intervention cycle to the [number missing]. One intervention cycle, This represents the preset number of evaluation cycles.

[0061] It should be noted that, The calculation method is as follows: Obtain the set of nodes of the direct-effect indicators of the intervention in the causal directed acyclic graph after the intervention, and calculate the node set of each indicator in the direct-effect indicator node set at the 1st... The difference between the actual observed value and the corresponding expected value in the counterfactual baseline sequence for each intervention period is weighted and summed according to the direct causal effect strength coefficient of each indicator on the target indicator. The result is the sum of these differences. . By from Subtract from the middle in turn , and The four components from each period are summarized to generate a four-component corrected attribution result.

[0062] It should be noted that the data quality interference component Separating the causal transmission effect into an independent component, the data quality interference component. This refers to the portion of the original attribution that is incorrectly included in the causal transmission effect due to data incompleteness asymmetry. Data quality interference component. The value reflects the degree to which differences in data quality interfere with attribution conclusions.

[0063] Step 8: Generate a comprehensive analysis report on the effectiveness of chronic disease quality control intervention.

[0064] The confidence intervals of each component of the four-component modified attribution results are calculated to assess the reliability of the attribution conclusions. Suggestions for improving data quality are summarized, and a comprehensive analysis report on the effectiveness of chronic disease quality control intervention is generated.

[0065] Step 8 may specifically include: Step S801: Calculate the confidence interval for each component using Bootstrap resampling on the four-component corrected attribution results. The input to Bootstrap resampling is the pre-intervention dataset and the post-intervention dataset after missing data processing, and the output is the upper and lower bounds of the confidence interval for each component estimate.

[0066] Step S802: Calculate the proportion of the data quality interference component in the total effect. When the proportion of the data quality interference component in the total effect exceeds a preset threshold, mark the attribution results with an insufficient attribution reliability warning.

[0067] Step S803: Summarize the indicator pairs corresponding to the difference edges with a probability of falsehood greater than zero in the causal structure difference list with missing risk labeling, the corresponding missing cause categories in the missing cause classification labeling results, and the institutional identifiers involved in management missingness as data quality improvement suggestions.

[0068] Step S804 integrates and outputs the corrected attribution estimate, confidence intervals for each component, reliability rating, and data quality improvement suggestions to generate a comprehensive analysis report on the effectiveness of chronic disease quality control intervention.

[0069] In this embodiment, to enable management to classify attribution reliability, the reliability rating is based on the proportion of data quality interference components to the total effect. When the proportion of data quality interference components to the total effect is lower than a first preset threshold, the reliability rating is high reliability; when the proportion of data quality interference components to the total effect is between the first preset threshold and a second preset threshold, the reliability rating is medium reliability; when the proportion of data quality interference components to the total effect exceeds the second preset threshold, the reliability rating is low reliability and an attribution reliability insufficient warning is triggered. The first and second preset thresholds can be set according to the management's decision-making requirements.

[0070] In this embodiment of the application, in order to further improve the pertinence of the data quality improvement suggestions, based on step S803, the institutions involved in the management deficiencies are sorted by the frequency of deficiencies and the number of deficiencies. The institutions with the highest frequency of deficiencies and the highest number of deficiencies are marked as priority improvement targets, and the details of the deficiencies corresponding to each institution are included in the data quality improvement suggestions, so that the management department can promote data management improvement in a targeted manner.

[0071] This embodiment uses association rule mining on the feature table of missing data patterns in phased data to classify missing events into systemic missing events, capability missing events, and managerial missing events according to their causes. A differentiated processing strategy is then adopted for each category: linear extrapolation is used to fill in systemic missing events, while capability and managerial missing events are excluded. This differentiated processing ensures that causal tests are performed based on reliable data in both the pre- and post-intervention periods, avoiding situations where simple uniform filling masks data management deficiencies or ignores missing events, leading to inconsistencies in causal test conditions before and after the intervention.

[0072] Based on this, this embodiment calculates the probability of missing edges causing spoofing for each difference edge. This transforms the determination of differences in causal structure before and after intervention from a binary judgment to a probabilistic assessment. In the attribution decomposition stage, discount weights are used... The causal edges involving the risk of missing data are corrected so that the spurious increase in the causal transmission effect component due to data quality differences is separated into an independent data quality interference component. Data quality interference components The stripping process enables attribution conclusions to distinguish between the actual causal structure changes caused by the intervention and the existing causal relationships newly detected after data refinement, thereby suppressing the interference of spurious structural changes on attribution conclusions under conditions of asymmetric data completeness.

[0073] Furthermore, this embodiment provides management departments with a quantitative reference for the reliability of attribution conclusions by assessing the proportion of data quality interference components in the total effect and outputting a reliability rating. It also generates data quality improvement suggestions by summarizing the indicator pairs and institutional details involved in missing false edges, giving a clear direction for subsequent improvements in data collection and management.

[0074] Application Examples See Figures 2-9 Based on the big data-based chronic disease quality control data analysis and processing method in the above embodiments, the following specific application examples are proposed: A provincial health administration department collected panel data on diabetes quality control indicators from 18 medical institutions within its jurisdiction during cycles A to B (6 cycles before intervention) and cycles B to C (6 cycles after intervention) in 20XX. In cycle B, a "Standardized Diabetes Management Special Intervention" was launched, with the target indicator being the HbA1c compliance rate. In the early stages of the intervention, the historical information system construction levels of each institution varied significantly, resulting in prominent data gaps. In the later stages, as the intervention measures were implemented, data management was strengthened, and data completeness improved significantly. A causal attribution analysis of the effectiveness of this intervention is now required.

[0075] The management department extracted quality control indicator data from 18 institutions across 6 cycles before and 6 cycles after the intervention from the data platform. This data covered 5 indicators: HbA1c target achievement rate, fasting blood glucose control rate, follow-up completion rate, medication adherence rate, and complication screening rate. The total number of institution cycles was 108 (before intervention) and 108 (after intervention). The system marked the completeness status of each indicator for each institution cycle and calculated the effective data coverage rate for each indicator in each period.

[0076] Table 1. Statistics on the coverage rate of effective data for phased quality control indicators Before the intervention, the effective coverage rates of various indicators were generally lower than after the intervention. The coverage rate of complication screening before the intervention was only 0.61, and the asymmetry of data completeness was the most prominent.

[0077] The system extracted all missing event records from the phased data missing pattern feature table, identifying a total of 142 missing events (136 before intervention and 6 after intervention). Using the institution level, information system version, and missing time range as itemsets, an association rule mining algorithm (minimum support set to 0.15, minimum confidence set to 0.70) was applied to extract association rules that met the threshold conditions, and the causes of each missing event were categorized and labeled.

[0078] Association rule mining results show that: for institutions using the V1.2 information system, the complication screening rate and follow-up completion rate were collectively missing in the 2nd to 3rd cycles before the intervention, with a support of 0.22 and a confidence of 0.81, which was judged as systematic missing; for institutions below level 3, the complication screening rate was continuously missing for 5 or more consecutive cycles before the intervention, with a support of 0.18 and a confidence of 0.76, which was judged as capability missing; the remaining scattered missing events did not meet the above two types of rules and were judged as management missing.

[0079] Table 2. Classification and labeling results of missing event causes (pre-intervention stage, some examples) Based on the above classification results, the system performs differentiated missing data processing. For systematic missing events such as M001 and M002, the system obtains the observation sequence of the most recent available period for the corresponding institution on the corresponding indicator, and calculates the filler value using linear extrapolation. For example, if institution 07's complication screening rate observation value is 0.38 in period 1, and is missing in periods 2 and 3 due to system upgrade interruption (V1.2), the filler values ​​extrapolated from the period 1 value are 0.40 and 0.42 respectively, and are written into the corresponding missing positions. For capability missing events such as M003 and management missing events such as M004 and M005, all indicator data of the corresponding institution in the corresponding period are marked as excluded and not included in subsequent calculations.

[0080] After processing, the effective sample size for each indicator in the early and late stages of intervention is calculated, and an effective sample size record table is generated.

[0081] Table 3 Key Indicators and Effective Sample Size Record Table The HbA1c target achievement rate → complication screening rate and follow-up completion rate → complication screening rate were both lower than the effective sample sizes of 51 and 47 before the intervention, respectively. The next step is to conduct a risk assessment of missing or falsified edges on these two different edges.

[0082] Based on the pre-intervention dataset after missing data processing, Granger causality tests (lag order 2) were applied to the time series data of the five indicators, outputting the causal direction determination for each indicator pair. The values ​​are then used, and conditional independence constraints are applied using the PC algorithm to generate a causal directed acyclic graph (DAG) before the intervention. The same operation is performed on the post-intervention dataset to generate a post-intervention causal DAG. Each causal edge is accompanied by... Value, causal effect strength coefficient And the corresponding effective sample size.

[0083] Table 4. Set of edges in the causal directed acyclic graph before and after intervention and confidence level labeling (edges related to HbA1c achievement rate). HbA1c target achievement rate → complication screening rate and follow-up completion rate → The complication screening rate failed the significance test before the intervention due to insufficient sample size, and only appeared in the post-intervention graph, where it was identified as a new causal edge.

[0084] The system compares the structural differences between the two graphs before and after the intervention, identifies two newly added causal edges, and calculates the probability of missing data in conjunction with the effective sample size record table.

[0085] Taking the HbA1c target achievement rate → complication screening rate as an example, this indicator reflects the effective sample size before intervention. , ,because The probability of missing data causing spurious data is: Follow-up completion rate → Complication screening rate: Effective sample size before intervention Similarly, .

[0086] in, The probability of missing data leading to falsification. The effective sample size before intervention. This is the minimum sample size threshold.

[0087] Table 5. List of causal structural differences with missing risk labels Although the follow-up completion rate → HbA1c target achievement rate represents a change in effect strength, the effective sample size before and after the intervention both meet the requirements. Require, This reflects the true effectiveness of the intervention.

[0088] Extract the set of parent nodes for HbA1c target achievement rate from the pre-intervention causal directed acyclic graph, including the follow-up completion rate ( ), fasting blood glucose control rate ( ) and medication standardization rate ( ), intercept term After normalizing the mean of the time series data of the three parent node indicators prior to intervention, linear extrapolation was used to calculate the expected values ​​of each indicator in the six periods following the intervention. Then, by weighted summation according to the formula, the counterfactual baseline sequence of HbA1c compliance rate is obtained.

[0089] in, For the first The counterfactual expected value of HbA1c compliance rate over a period of time. For the first The causal effect strength coefficient of each parent node, For the first The parent node index is at the 1st Extrapolated expected value of the cycle This is the intercept term. Taking the first cycle after intervention as an example, the extrapolated expected values ​​of the three parent node indicators are 0.71, 0.68, and 0.74, respectively. Substituting these values ​​into the formula, we get: Table 6 Comparison of HbA1c compliance rate counterfactual baseline sequence and actual value The fact that the actual observed values ​​were consistently higher than the counterfactual expected values ​​indicates that the intervention measures had a positive effect on the HbA1c target achievement rate.

[0090] The system first discounts and corrects causal edges along the causal transmission path that involve the risk of omissions or falsification. Taking the third cycle after intervention as an example, in the path from follow-up completion rate to complication screening rate to HbA1c target achievement rate, the percentage difference between follow-up completion rate and complication screening rate is... Original causal effect intensity coefficient After correction: in, This is the corrected causal effect strength coefficient. This is the original causal effect strength coefficient. This represents the probability of spurious errors due to the missing causal edge. After correction, the causal transmission effect component is correspondingly reduced, and the discount difference is separated into a data quality interference component. For the third period after intervention ( The four-component decomposition results are as follows: Table 7. Four-component adjusted attribution results (each cycle after intervention) The data quality interference component remained at a low level in each period, accounting for approximately 5% to 6% of the total difference, indicating that the risk of false positives due to missing data had a limited impact on the attribution conclusions.

[0091] The system performs Bootstrap resampling (1000 resampling times) on the four-component corrected attribution results and calculates the 95% confidence interval for each component. The average proportion of data quality interference components in the total effect is 5.3%, which is lower than the first preset threshold (15%), and the reliability rating is high reliability, so there is no need to trigger the attribution reliability insufficient warning. The system also summarizes the difference edges with a false probability greater than zero (HbA1c compliance rate → complication screening rate, follow-up completion rate → complication screening rate), and combined with the classification and labeling of the causes of the missing data, the institutions 07 and 11 using the V1.2 system are marked as priority rectification targets for systemic missing data, the institution 04 involved in the capability missing data is included in the capability improvement suggestion list, and the institutions 13 and 09 with the highest frequency of management missing data are marked as priority improvement targets for data management. Finally, the system integrates and outputs a comprehensive analysis report on the effectiveness of chronic disease quality control intervention.

[0092] The entire data flow process reflects a complete logical chain from raw panel data to the final attribution report: the coverage statistics in step 1 provide input for the association rule mining in step 2; the classification labeling in step 2 determines the differentiation processing strategy in step 3; the effective sample size record table output in step 3 directly supports the calculation of the probability of missing data in step 5; the causal graph in step 4 and the risk labeling in step 5 together provide the basis for the counterfactual extrapolation in step 6 and the discount correction in step 7; the four-component decomposition results in step 7 are finally incorporated into the comprehensive report in step 8, ensuring that the attribution conclusions remain reliable under the condition of asymmetric data completeness.

[0093] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A big data-based chronic disease quality control data analysis and processing method, characterized in that, include: Acquire panel data of quality control indicators from multiple medical institutions before and after the intervention, divide them into pre-intervention datasets and post-intervention datasets according to the intervention time nodes, and statistically analyze the effective data coverage and missing pattern characteristics of each indicator in each period to generate a staged quality control indicator panel dataset and a staged data missing pattern characteristic table. For the phased data missing pattern feature table, the association rule mining algorithm is used to analyze the relationship between missing events and institutional attributes and time factors, and each missing event is classified into three cause categories: systemic missing, capability missing and management missing, and the missing cause classification labeling results are generated. Based on the classification and labeling results of the missing causes, differential missing data processing is performed on the pre-intervention dataset and the post-intervention dataset. Systematic missing data is filled using linear extrapolation, while capability missing data and management missing data are marked as excluded. A staged dataset and an effective sample size record table are generated after missing data processing. Causal tests are applied to each stage dataset after missing data processing to generate causal directed acyclic graphs before and after intervention. Structural differences are compared between the causal directed acyclic graphs before and after intervention. Based on the effective sample size record table, the probability of missing data causing false positives is calculated for each difference edge, and a list of causal structural differences with missing data causing false positives is generated. Based on the causal structure of the pre-intervention causal directed acyclic graph and the time trend of the pre-intervention data, counterfactual extrapolation is performed on the target indicators to generate a counterfactual baseline sequence. The difference sequence between the actual value sequence after intervention of the target indicator and the counterfactual baseline sequence is calculated. Causal edges involving missing or falsified risks on the causal path are corrected by discount weights. The difference sequence is decomposed into direct intervention effect component, corrected causal transmission effect component, data quality interference component, and external factor residual component to generate a four-component corrected attribution result.

2. The big data-based chronic disease quality control data analysis processing method according to claim 1, characterized in that, The algorithm for analyzing the association between missing events and organizational attributes and time factors using association rule mining categorizes missing events into three causes: systemic missing events, capability missing events, and managerial missing events. Extract missing event records from the missing pattern feature table of the phased data. Each missing event record contains a missing indicator identifier, a missing organization identifier, a missing cycle identifier, and corresponding organization attribute features. The organization attribute features include the organization level, the region, and the information system version. Using missing event records as transaction sets and organizational attributes and time factors as itemsets, an association rule mining algorithm is applied to extract a set of association rules that satisfy the preset minimum support and minimum confidence. When the association rule corresponding to a missing event points to a batch of missing information system versions within a specific time period, the missing event is marked as a systemic missing event; when the association rule corresponding to a missing event points to a persistent missing indicator for an organization at a specific organizational level, the missing event is marked as a capability missing event; when a missing event does not meet the matching conditions of the above two types of association rules, the missing event is marked as a management missing event. 3.The big data-based chronic disease quality control data analysis processing method according to claim 1, characterized in that, The method of filling systematic missing information using linear extrapolation includes: For missing events marked as systematically missing, obtain the sequence of observations for the most recent available period for the corresponding institution on the corresponding indicator, calculate the filling value for the missing period using linear extrapolation, and write the filling value into the missing position in the corresponding dataset; For missing events marked as capability-related or managerial-related deficiencies, no filling process is performed; instead, all indicator data of the corresponding institution in that period are marked as excluded. The effective sample size record table records the number of effective observations for each pair of indicators in the early and late stages of the intervention.

4. The method for analyzing and processing chronic disease quality control data based on big data according to claim 1, characterized in that, The process of applying causal tests to each stage dataset after missing data processing to generate causal directed acyclic graphs before and after intervention includes: For each stage dataset after missing data processing, Granger causality test based on vector autoregression model is applied to generate candidate causal edge set, and then PC algorithm is applied to generate causal directed acyclic graph after conditional independence test constraint. For each causal edge, record its statistical test p-value, causal effect strength coefficient, and corresponding effective sample size. The causal effect strength coefficient is the regression coefficient of the corresponding lag term in the vector autoregression model. Append the above information to each edge of the causal directed acyclic graph to generate causal directed acyclic graphs before and after intervention with test confidence level labels.

5. The method for analyzing and processing chronic disease quality control data based on big data according to claim 1, characterized in that, The calculation of the missing probability for each dissimilar edge based on the effective sample size record table includes: Edges that exist only in the post-intervention causal directed acyclic graph are marked as newly added causal edges; edges that exist only in the pre-intervention causal directed acyclic graph are marked as disappeared causal edges; edges that exist in both graphs but whose causal effect intensity coefficient changes by more than a preset change threshold are marked as effect intensity change edges. For a newly added causal edge, obtain the effective sample size of the corresponding indicator pair in the early stage of intervention and the minimum sample size threshold required for causal testing. When the effective sample size in the early stage of intervention is less than the minimum sample size threshold, the probability of missing data being false is the ratio of the difference between the minimum sample size threshold and the effective sample size in the early stage of intervention divided by the minimum sample size threshold. When the effective sample size in the early stage of intervention is not less than the minimum sample size threshold, the probability of missing data being false is zero. For the missing causal edge, obtain the effective sample size of the corresponding indicator pair in the later stage of the intervention, and calculate the probability of missing spurious edges in the same way as the newly added causal edge. For the edge of effect intensity change, when the effective sample size in either the early or late stage of the intervention is lower than the minimum sample size threshold, the probability of missing data is calculated in the above manner; when the effective sample size in both stages is not lower than the minimum sample size threshold, the probability of missing data is zero.

6. The method for analyzing and processing chronic disease quality control data based on big data according to claim 1, characterized in that, The method involves counterfactual extrapolation of the target indicators based on the causal structure of the pre-intervention causal directed acyclic graph and the time trend of pre-intervention data, generating a counterfactual baseline sequence, including: Extract the set of parent nodes of the target indicator and the corresponding causal effect strength coefficient from the pre-intervention causal directed acyclic graph; The extrapolated expected values ​​of each parent node indicator in each period after intervention were calculated using a linear extrapolation method. The extrapolated expected values ​​of each parent node indicator are weighted and summed with the corresponding causal effect strength coefficients. The intercept term obtained by fitting the data before the intervention is added to calculate the counterfactual expected values ​​of the target indicator in each period after the intervention. The counterfactual expected values ​​of each period are then summarized to generate a counterfactual baseline sequence. In this process, the time series of each parent node index is normalized to mean based on the range before linear extrapolation, and the causal effect intensity coefficient is estimated based on the normalized data.

7. The method for analyzing and processing chronic disease quality control data based on big data according to claim 1, characterized in that, The method of correcting causal edges on causal paths that involve the risk of missing information or falsification by applying discount weights includes: Based on the causal path from the directly affected indicator to the target indicator in the post-intervention causal directed acyclic graph, the original components of the causal transmission effect are calculated. For each causal edge involved in the causal path, query the list of causal structure differences marked with missing and falsification risks. For causal edges marked as missing and falsification, multiply the original causal effect strength coefficient by one and subtract the discount weight obtained from the corresponding missing and falsification probability to obtain the corrected causal effect strength coefficient. The causal transmission effect component is recalculated based on the corrected causal effect intensity coefficient to generate the corrected causal transmission effect component. The data quality interference component is the difference between the original causal transmission effect component and the corrected causal transmission effect component, and the external factor residual component is the remaining difference in the difference sequence after deducting the direct intervention effect component, the corrected causal transmission effect component, and the data quality interference component.

8. The method for analyzing and processing chronic disease quality control data based on big data according to claim 7, characterized in that, The direct intervention effect component is calculated as follows: obtain the node set of the direct intervention indicators in the causal directed acyclic graph after intervention, and sum the differences between the actual observed values ​​of each indicator in the node set of direct intervention indicators and the corresponding expected values ​​in the counterfactual baseline sequence in each intervention period, according to the direct causal effect intensity coefficient of each indicator on the target indicator.

9. The method for analyzing and processing chronic disease quality control data based on big data according to claim 1, characterized in that, Also includes: The confidence intervals of each component are calculated using Bootstrap resampling on the four-component corrected attribution results; The proportion of the data quality interference component to the total effect is calculated, and a reliability rating is performed based on the proportion. When the proportion is lower than the first preset threshold, it is rated as high reliability. When the proportion is between the first preset threshold and the second preset threshold, it is rated as medium reliability. When the proportion exceeds the second preset threshold, it is rated as low reliability and an attribution reliability insufficient warning is triggered. The indicator pairs corresponding to the difference edges with a probability of missing falsehood greater than zero in the causal structure difference list with missing falsehood risk labeling, the corresponding missing cause categories in the missing cause classification labeling results, and the institutional identifiers involved in management missingness are summarized as data quality improvement suggestions. The corrected attribution estimate, confidence intervals of each component, reliability rating and data quality improvement suggestions are integrated and output to generate a comprehensive analysis report on the effect of chronic disease quality control intervention.

10. A big data-based chronic disease quality control data analysis and processing system, used to execute the big data-based chronic disease quality control data analysis and processing method according to any one of claims 1 to 9, characterized in that, include: The phased data acquisition module is used to acquire panel data of quality control indicators from multiple medical institutions before and after the intervention. It is divided into pre-intervention datasets and post-intervention datasets according to the intervention time nodes, and the effective data coverage and missing pattern characteristics of each indicator in each period are statistically analyzed to generate phased quality control indicator panel datasets and phased data missing pattern characteristic tables. The missing cause classification module is used to analyze the relationship between missing events and organizational attributes and time factors in the missing pattern feature table of the phased data using association rule mining algorithm, classify each missing event into three cause categories: systemic missing, capability missing and management missing, and generate missing cause classification labeling results; The differential missing data processing module is used to perform differential missing data processing on the pre-intervention dataset and the post-intervention dataset based on the missing data cause classification and labeling results. Systematic missing data is filled using linear extrapolation, while capability missing data and management missing data are marked as excluded. The module generates a staged dataset and an effective sample size record table after missing data processing. The causal structure difference analysis module is used to apply causal tests to the datasets of each stage after missing data processing to generate causal directed acyclic graphs before and after intervention. The module compares the structural differences between the causal directed acyclic graphs before and after intervention, calculates the probability of missing data causing false positives for each difference edge based on the effective sample size record table, and generates a list of causal structure differences with missing data causing false positives labeling. The counterfactual extrapolation module is used to perform counterfactual extrapolation on the target indicators based on the causal structure of the pre-intervention causal directed acyclic graph and the time trend of the pre-intervention data, and generate a counterfactual baseline sequence. The modified attribution decomposition module is used to calculate the difference sequence between the actual value sequence after intervention of the target indicator and the counterfactual baseline sequence. It modifies the causal edges involving the risk of missing data and falsification on the causal path according to the discount weight, and decomposes the difference sequence into the direct intervention effect component, the modified causal transmission effect component, the data quality interference component, and the external factor residual component, generating a four-component modified attribution result.

Images