An optimization method, device and medium of a curative effect evaluation index system

Abstract

The application discloses a kind of curative effect evaluation index system optimization method, equipment and medium, it is related to index optimization technical field, including: obtaining the analysis data set after merging;Based on multivariate statistical method and machine learning algorithm, the feature importance analysis and regression fitting are carried out to the data set, the influence coefficient and correlation weight of each index to curative effect result are recalculated, and the index weight after optimization is obtained;Based on the index weight after optimization, the weight value in the original curative effect evaluation index system is replaced, and the curative effect evaluation index system after optimization is generated.The application fuses retrospective clinical data set and prospective randomized controlled test data set, and based on multivariate statistical method and machine learning algorithm, the influence coefficient and correlation weight of each index are recalculated, the weight of the original curative effect evaluation index system is replaced, realizes the objective verification and scientific and technological standardization of the multidimensional curative effect evaluation index system that has been initially established, and improves the scientificity and accuracy of system.

Description

Technical Field

[0001] This invention relates to the field of indicator optimization technology, and in particular to a method, equipment and medium for optimizing a therapeutic efficacy evaluation indicator system. Background Technology

[0002] After establishing a multidimensional efficacy evaluation index system for obese polycystic ovary syndrome (PCOS), the clinical applicability and evaluation accuracy of this system often still require further verification and calibration. The initially constructed evaluation system is usually based on historical data such as retrospective electronic medical records. While this type of data has certain advantages in sample size, the inconsistencies in the completeness of medical records and the fact that patient grouping is not randomized mean that the weights of the indicators derived from retrospective data may deviate from the actual clinical intervention effects, thus affecting the accuracy of the evaluation system in reflecting the effectiveness and safety of drugs or treatment regimens.

[0003] In existing technologies, to verify and improve the reliability of evaluation systems, researchers conduct prospective randomized controlled trials to obtain higher-quality prospective clinical data through rigorous trial design and standardized data collection. The inventors of this application have found that existing system optimization methods primarily rely on fine-tuning parameters using a single mathematical model or statistical indicator. High-quality data obtained from prospective trials are often only used for independent efficacy comparison analyses. After the evaluation system is constructed using preliminary retrospective data, the high-quality prospective data collected again is difficult to systematically apply to the weight calibration and iterative improvement of the already constructed system. This causes the evaluation system to become rigid once initially established, unable to continuously improve its accuracy and scientific rigor with the accumulation of new evidence. Summary of the Invention

[0004] This invention provides a method, device, and medium for optimizing a therapeutic efficacy evaluation index system. The technical problem it aims to solve is: how to provide an effective solution for objectively verifying and scientifically calibrating a preliminarily established multi-dimensional therapeutic efficacy evaluation index system.

[0005] In a first aspect, the present invention provides a method for optimizing a therapeutic efficacy evaluation index system, comprising: Obtain the merged analysis dataset, which is obtained by merging retrospective clinical datasets and prospective randomized controlled trial datasets. The merged analysis dataset contains clinical indicator data related to multiple evaluation dimensions recorded under multiple treatment modalities. Based on multivariate statistical methods and machine learning algorithms, feature importance analysis and regression fitting are performed on the merged analysis dataset. The influence coefficients and correlation weights of each indicator on the efficacy results are recalculated to obtain the optimized indicator weights. Based on the optimized index weights, the weight values ​​in the original efficacy evaluation index system are replaced to generate the optimized efficacy evaluation index system.

[0006] Optionally, the retrospective clinical dataset and the prospective randomized controlled trial dataset are merged to obtain the merged analysis dataset, which includes: Obtain the retrospective clinical dataset and the prospective randomized controlled trial dataset; Field mapping is performed on the retrospective clinical dataset and the prospective randomized controlled trial dataset to uniformly map field names with the same clinical meaning to preset standardized fields; The mapped data is normalized to eliminate the dimensional differences between the retrospective clinical dataset and the prospective randomized controlled trial dataset; The normalized data is cleaned and verified to obtain the merged analysis dataset.

[0007] Optionally, the step of performing feature importance analysis and regression fitting on the merged analysis dataset based on multivariate statistical methods and machine learning algorithms, recalculating the influence coefficients and correlation weights of each indicator on the efficacy results, and obtaining optimized indicator weights includes: The first feature importance analysis was performed on the merged analysis dataset based on multivariate statistical methods to obtain the first importance score of each indicator; The merged analysis dataset is subjected to a second feature importance analysis based on machine learning algorithms to obtain the second importance score of each indicator; The first importance score and the second importance score are normalized respectively to obtain the normalized first importance score and the normalized second importance score; The weighted sum of the normalized first importance score and the normalized second importance score is used to obtain the fused importance score of each indicator. The fusion importance score is used as the influence coefficient and correlation weight of each indicator on the therapeutic effect, and is used as the optimized indicator weight.

[0008] Optionally, the multivariate statistical method includes a generalized estimation equation, and the first feature importance analysis is performed on the merged analysis dataset based on the multivariate statistical method to obtain the first importance score of each indicator, including: Obtain repeated measurement data of the same patient at different time points in the merged analysis dataset; A generalized estimation equation model was constructed, with treatment method and measurement time as independent variables and the changes in each clinical indicator as dependent variables, and the job correlation matrix between the repeated measurement data was set. The effect estimates of each independent variable are calculated using the generalized estimation equation model, and the absolute values ​​of the effect estimates are used as the first importance scores for each indicator.

[0009] Optionally, setting the job correlation matrix between the repeated measurement data includes: The measurement time interval for obtaining the repeated measurement data; If the measurement time intervals are not equal, then select the unstructured operation correlation matrix; If the measurement time intervals are equal, the correlation coefficient between each measurement time point is calculated, and an autoregressive job correlation matrix or a commutative job correlation matrix is ​​selected based on the changing trend of the correlation coefficient between each measurement time point with the measurement time interval.

[0010] Optionally, the machine learning algorithm includes random forest, and the second feature importance analysis is performed on the merged analysis dataset based on the machine learning algorithm to obtain the second importance score of each indicator, including: Multiple bootstrap samples are extracted from the merged analysis dataset. A classification and regression decision tree is constructed based on each bootstrap sample. When splitting at each node of the classification and regression decision tree, a subset of features is randomly selected from the feature set to select the optimal splitting feature, thereby constructing a random forest model. The average reduction in impurity of each feature in the random forest model across all decision trees is calculated and used as the second importance score for each indicator.

[0011] Optionally, calculating the average reduction in impurity of each feature in the random forest model across all decision trees, as the second importance score for each metric, includes: When constructing the random forest model, for each decision tree, the prediction error of the decision tree is calculated using the out-of-bag data that was not selected in the sampling. The feature values ​​of a certain feature in the out-of-bag data are randomly shuffled, and the prediction error of the decision tree after shuffling is calculated again. Calculate the increase in prediction error before and after shuffling, and calculate the average of the increase across all decision trees. Use the average as the second importance score for each indicator.

[0012] Optionally, the step of replacing the weight values ​​in the original efficacy evaluation index system based on the optimized index weights to generate the optimized efficacy evaluation index system includes: Obtain the original efficacy evaluation index system, which includes preset evaluation dimensions, the original weight values ​​of each index item, and scoring rules; Replace the original weight values ​​with the optimized index weights; Based on the replaced indicator weights, the scoring rules are updated to generate an optimized efficacy evaluation indicator system.

[0013] Secondly, the present invention also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0014] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0015] This invention provides a method, device, and medium for optimizing a efficacy evaluation index system. By acquiring a merged analysis dataset obtained from retrospective clinical datasets and prospective randomized controlled trial datasets, the recalculation of index weights can be based on data from two different sources and levels of evidence—retrospective clinical data and prospective randomized controlled trial data—overcoming the limitation of the initially established multidimensional efficacy evaluation index system relying solely on a single retrospective data source and lacking objective verification and scientific calibration based on prospective data. Based on this, feature importance analysis and regression fitting are performed on the merged analysis dataset using multivariate statistical methods and machine learning algorithms. The correlation between each index and efficacy results is quantitatively assessed and recalculated using two different analytical paradigms to obtain optimized index weights. The redefined weight parameters have higher reliability compared to results obtained using a single assessment method. Furthermore, the optimized index weights are used to replace the weight values ​​in the original efficacy evaluation index system, generating an optimized efficacy evaluation index system. This allows the system to obtain weight parameter corrections adapted to prospective clinical evidence while retaining the original evaluation framework, thereby achieving objective verification and scientific calibration of the initially established multidimensional efficacy evaluation index system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for optimizing an efficacy evaluation index system provided in an embodiment of the present invention; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [the described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [the described condition or event] is detected," or "in response to detection of [the described condition or event]."

[0023] In the preliminary work, the applicant has constructed a "disease-syndrome-symptom-detection indicator" efficacy evaluation system through statistical screening of retrospective clinical data and expert consultation. This system includes four evaluation dimensions: disease indicators, syndrome dimensions, symptom dimensions, and detection indicator dimensions. Each dimension contains several specific indicators, each assigned a corresponding weight and a scoring rule established. However, the indicator weights in this system are primarily derived from historical medical record data. Limited by the inherent selection bias, information bias, and confounding bias of retrospective data, this may lead to discrepancies between the weight settings and the actual prospective clinical intervention effects. It should be noted that confounding factors refer to external variables that are simultaneously associated with treatment selection and efficacy outcomes, potentially leading to incorrect attribution of efficacy differences. In evidence-based medicine, prospective randomized controlled trials are recognized as a high-level evidence method for evaluating the effectiveness of interventions. Through randomization, prospective follow-up, and standardized data collection, they can effectively control for known and unknown confounding factors, leading to more reliable causal inferences. Therefore, after completing the prospective randomized controlled trial and obtaining relevant data, it is necessary to use this high-quality data to validate and recalibrate the initially established indicator system.

[0024] Therefore, embodiments of the present invention provide a method for optimizing a therapeutic efficacy evaluation index system, particularly an optimization method for prospectively validating and dynamically calibrating a preliminarily established multi-dimensional therapeutic efficacy evaluation index system for obese polycystic ovary syndrome. Please refer to [link to relevant documentation]. Figure 1 The method includes the following steps: S1, Obtain the merged analysis dataset, which is obtained by merging retrospective clinical datasets and prospective randomized controlled trial datasets. The merged analysis dataset contains clinical indicator data related to multiple evaluation dimensions recorded under various treatment methods.

[0025] In practice, the retrospective clinical dataset is derived from the in-hospital electronic medical record data used in the initial construction of the efficacy evaluation index system. This data is naturally accumulated during routine diagnosis and treatment and falls under the category of real-world data. Real-world data is data collected in routine clinical practice that reflects the actual diagnosis and treatment process and health status of patients. Its advantages lie in its typically large sample size, wide coverage of patient types, and ability to reflect treatment effects in a real clinical environment. However, its limitations include inconsistent data record completeness and standardization, and the fact that patient grouping is not random, with potential systematic differences in baseline characteristics between different treatment groups, which can contaminate treatment effects and interfere with the objectivity of efficacy evaluation.

[0026] The prospective randomized controlled trial dataset is derived from subsequent clinical trials that strictly adhere to a randomized controlled design. A randomized controlled trial (RCT) is a research method that randomly assigns research subjects to different intervention and control groups, and conducts simultaneous follow-up observations under identical conditions to evaluate the intervention's effectiveness. Its core lies in randomization, ensuring that each patient has an equal opportunity to enter either treatment group, thus evenly distributing known and unknown confounding factors across groups and ensuring that differences in efficacy between groups are attributable to the treatment intervention itself. The prospective trial can set up a treatment group (receiving a combined traditional Chinese and Western medicine treatment plan) and a control group (receiving a standard Western medicine treatment plan). Various clinical indicators are collected at multiple time points before, during, and after treatment, following a standardized procedure. Multiple treatment methods refer to the aforementioned combined traditional Chinese and Western medicine plan and the standard Western medicine plan. The multiple evaluation dimensions in this invention include disease indicator dimensions, syndrome dimensions, symptom dimensions, and detection indicator dimensions. The clinical indicator data include menstrual status, weight, body mass index, waist circumference, waist-to-hip ratio, sex hormone levels, insulin resistance index, glucose tolerance, blood lipid biochemical indicators, uric acid, and fatty liver indicators.

[0027] It should be noted that when acquiring the above two types of datasets, the present invention has anonymized the data and removed the patient's personal identity information to protect the patient's privacy.

[0028] In some preferred embodiments, the retrospective clinical dataset and the prospective randomized controlled trial dataset are merged to obtain a merged analysis dataset, including: acquiring the retrospective clinical dataset and the prospective randomized controlled trial dataset; performing field mapping on the retrospective clinical dataset and the prospective randomized controlled trial dataset, uniformly mapping field names with the same clinical meaning to preset standardized fields; normalizing the mapped data to eliminate dimensional differences between the retrospective clinical dataset and the prospective randomized controlled trial dataset; and cleaning and validating the normalized data to obtain the merged analysis dataset.

[0029] In practice, retrospective clinical data originates from the natural accumulation of daily medical records, while prospective randomized controlled trial data comes from specially designed clinical trials. These two types of data typically differ significantly in terms of data collection objectives, data recording standards, field naming systems, coding rules, and units of measurement. For example, weight in retrospective medical records might be described in "jin" (a unit of weight in kilograms) in the medical record text, while prospective trial data is entered in "kilograms" using a standardized case report form. This multi-source heterogeneity means that the two types of data cannot be directly merged and must be integrated through a system. The processing flow in this embodiment is as follows: We acquired retrospective clinical datasets (structured data extracted from multiple heterogeneous data sources within the hospital through extraction, entity recognition, relation extraction, and standardized mapping) and prospective randomized controlled trial datasets (collected during concurrent clinical trials, with a combination of traditional Chinese and Western medicine treatment group and a conventional Western medicine control group).

[0030] Next, field mapping is performed on these two datasets, uniformly mapping clinically meaningful field names to pre-defined standardized fields. Field mapping aims to establish semantic correspondences between fields. Because data collected by different teams or systems may use different names even for the same indicator—for example, "luteinizing hormone" in a retrospective database might be written as "luteinizing hormone" or an abbreviation in a prospective database. Therefore, by pre-constructing a synonymous field mapping dictionary, fields with different names but the same meaning can be uniformly mapped to the same standardized name, ensuring correct identification and correspondence. After mapping, an intermediate dataset with unified field names is obtained.

[0031] Then, the mapped data is normalized to eliminate the dimensional differences between the two datasets. Normalization is an operation that unifies numerical values ​​of different dimensions or orders of magnitude to the same scale through mathematical transformation. Common methods include linear normalization (mapping to the [0,1] interval) and standard deviation normalization (converting to standard scores with a mean of 0 and a standard deviation of 1). This embodiment mainly addresses the problem of inconsistent units of measurement, such as converting weight units from "jin" to "kilogram" and hormone levels from traditional units to international standard units. After normalization, the problem of data incomparability caused by inconsistent units is solved, ensuring that the numerical scales are consistent and the orders of magnitude are comparable in subsequent analyses.

[0032] Finally, the normalized data was cleaned and validated to obtain the merged analysis dataset. Cleaning and validation involves a comprehensive check and correction of data quality, including: identifying and removing duplicate records caused by entry or export errors; employing appropriate strategies for missing values ​​based on their proportion and importance (e.g., filling in continuous variables with low missing proportions using the mean or median, and removing those with high missing proportions or missing key fields); identifying records with significantly abnormal values ​​(e.g., negative weight or weight exceeding physiological limits), marking them as missing or truncating them; and verifying data consistency, such as whether there are logical contradictions in the correlation indicators of the same patient at the same time point, marking them and manually reviewing or removing them if any are found. After field mapping, normalization, and cleaning and validation, the two datasets, which originally differed significantly in format, units, and quality, were integrated into a merged analysis dataset with a unified format, consistent dimensions, and controllable quality.

[0033] This embodiment organically integrates retrospective and prospective datasets into a unified and usable analytical dataset through a standardized integration process of field mapping, normalization, and cleaning and verification. Field mapping, utilizing a synonymous field mapping dictionary, solves the difficulty in field identification caused by different naming conventions, preventing data loss or incorrect merging. Normalization, by unifying units, eliminates numerical incomparability caused by unit differences, ensuring fair treatment of the same indicator across different sources and preventing bias in weight calculation. Cleaning and verification, through multi-dimensional quality checks and corrections, eliminates interference from outliers and erroneous data, improving the intrinsic quality of the analytical dataset and the reliability of weight recalculation. The effective integration of these two types of data gives the analytical dataset the advantages of both the large sample size and broad coverage of retrospective data and the standardized collection and strong causal inference capabilities of prospective data, providing a superior foundation for subsequent multivariate statistical and machine learning weight recalculation in terms of data breadth, quality, and strength of evidence compared to a single data source.

[0034] S2, based on multivariate statistical methods and machine learning algorithms, feature importance analysis and regression fitting are performed on the merged analysis dataset, and the influence coefficients and correlation weights of each indicator on the efficacy results are recalculated to obtain the optimized indicator weights.

[0035] In practical implementation, multivariate statistical methods refer to statistical methods that can simultaneously analyze the relationships between multiple variables. They have clear probability distribution assumptions and statistical inference frameworks, and can provide parameter estimates, standard errors, and significance test results. Machine learning algorithms, on the other hand, automatically learn patterns from data to complete prediction or classification tasks. They do not rely on pre-set linear and distribution assumptions and can adaptively discover complex nonlinear relationships and variable interaction effects. Feature importance analysis is the process of evaluating the contribution of each input feature to the output result, with the output being an importance score for each feature; the higher the score, the greater the influence. Regression fitting establishes a mathematical function relationship between independent and dependent variables to approximate the data generation mechanism. The core of this step lies in using these two types of methods, which are fundamentally different in theoretical foundation and analytical approach, in parallel, and then fusing their results. Different algorithms make different assumptions about the data structure. A single algorithm may produce systematic biases due to mismatches between its assumptions and the actual data characteristics; however, the parallel application and fusion of multiple algorithms can achieve multi-perspective complementarity and cross-validation, making the final index weights more robust and reliable.

[0036] In some preferred embodiments, the step of performing feature importance analysis and regression fitting on the merged analysis dataset based on multivariate statistical methods and machine learning algorithms to recalculate the influence coefficients and correlation weights of each indicator on the efficacy results, and obtaining optimized indicator weights, includes: performing a first feature importance analysis on the merged analysis dataset based on multivariate statistical methods to obtain a first importance score for each indicator; performing a second feature importance analysis on the merged analysis dataset based on machine learning algorithms to obtain a second importance score for each indicator; normalizing the first importance score and the second importance score respectively to obtain a normalized first importance score and a normalized second importance score; weighting and summing the normalized first importance score and the normalized second importance score to obtain a fused importance score for each indicator; and using the fused importance score as the influence coefficient and correlation weight of each indicator on the efficacy results, as the optimized indicator weights.

[0037] In practice, after obtaining the merged analysis dataset, the core technical challenge of this step is how to comprehensively utilize multiple algorithms with fundamentally different analytical principles and mathematical foundations to objectively assess the relative importance of each clinical indicator to the therapeutic effect. If only a single algorithm is used, its conclusions may be limited by the algorithm's assumptions and analytical framework, making it difficult to fully reflect the complex relationships that may exist simultaneously in the data, such as linear associations, nonlinear associations, main effects, and interaction effects. Therefore, this embodiment achieves the systematic integration of results from multiple algorithms by performing multivariate statistical analysis and machine learning analysis in parallel, followed by normalization and weighted fusion of the results, as detailed below: First feature importance analysis was performed on the merged dataset using multivariate statistical methods to obtain the first importance score for each indicator. Statistical modeling and parameter estimation were then conducted on the differences in changes of each clinical indicator under different treatment methods using generalized estimation equations or analysis of variance. The effect size of each indicator on the difference in efficacy was calculated, and the absolute value or standardized value of the effect size was used as the first importance score for that indicator. The effect size is a standardized measure reflecting the magnitude of the influence of the independent variable on the dependent variable; its value directly represents the average magnitude of the change in the dependent variable when the independent variable changes by one unit.

[0038] Furthermore, a second feature importance analysis is performed on the merged analysis dataset based on machine learning algorithms to obtain the second importance score for each indicator. Algorithms such as random forest or principal component analysis are used to evaluate the relative importance of each clinical indicator in predicting efficacy outcomes from a data-driven perspective, serving as the second importance score for each indicator.

[0039] The first and second importance scores are normalized separately to obtain normalized first and second importance scores. The first importance scores output by multivariate statistical methods and the second importance scores output by machine learning algorithms may differ significantly in numerical dimensions and distribution range. For example, the effect estimates output by generalized estimation equations may be expressed in the original units of measurement, and their numerical range fluctuates greatly depending on the indicators; while the importance scores output by random forests are usually relative values, and their absolute values ​​depend on the model parameters and the size of the dataset. If two sets of scores with different numerical ranges and dimensions are directly merged, the set with the larger dimension will have a disproportionate impact on the fusion result, essentially giving that set a higher implicit weight. Normalization transforms the two sets of importance scores to a unified value range through mathematical transformation, making the two sets of scores directly comparable at the same scale.

[0040] The weighted sum of the normalized first importance score and the normalized second importance score yields the fusion importance score for each indicator. The weighting coefficients reflect the relative reliability of the results of multivariate statistical and machine learning analysis and can be flexibly configured according to the actual application scenario and data characteristics. This fusion importance score serves as the influence coefficient and correlation weight of each indicator on the efficacy outcome, i.e., the optimized indicator weights.

[0041] This embodiment achieves comprehensive evaluation and result integration of the same dataset from different analytical perspectives by performing multivariate statistical analysis and machine learning analysis in parallel and then normalizing and weighting the results. Multivariate statistical methods are based on rigorous statistical theory, and their analytical results have good interpretability and statistical inference support, providing quantitative evidence with clear statistical meaning for the importance of indicators. Machine learning algorithms do not rely on linear or distributional assumptions and have the ability to adaptively discover complex patterns in data, revealing nonlinear relationships and interaction effects that are difficult to capture by traditional statistical methods. Both methods independently evaluate the same set of data from different analytical paradigms, forming complementary analytical perspectives and effectively avoiding evaluation biases caused by the methodological limitations of a single analytical paradigm. Normalization eliminates the incomparability caused by differences in units and numerical ranges between the two analytical results, ensuring the fairness and rationality of the weighted fusion. Furthermore, the weighted fusion mechanism provides flexible space for adjusting the analytical emphasis, allowing the method to adapt the contribution ratio of the two analytical paradigms to the data characteristics and needs of different application scenarios, enhancing the applicability and customizability of the overall technical solution.

[0042] In some preferred embodiments, the multivariate statistical method includes a generalized estimation equation. The step of performing a first feature importance analysis on the merged analysis dataset based on the multivariate statistical method to obtain the first importance score for each indicator includes: obtaining repeated measurement data of the same patient at different time points in the merged analysis dataset; constructing a generalized estimation equation model, using treatment method and measurement time as independent variables, and the change in each clinical indicator as the dependent variable, and setting an operation correlation matrix between the repeated measurement data; calculating the effect estimate of each independent variable using the generalized estimation equation model, and using the absolute value of the effect estimate as the first importance score for each indicator.

[0043] In clinical studies, patients typically undergo repeated measurements and data collection at multiple time points, including baseline before treatment, multiple follow-up points during treatment, and after treatment. For example, in a 12-week treatment cycle, patients' weight, sex hormone levels, and other indicators are measured at baseline, week 4, week 8, and week 12. The data generated by this longitudinal study design exhibits a hierarchical structure; multiple measurements of the same patient at different time points form a data cluster, and there are varying degrees of correlation between measurements within a cluster. Typically, measurements at adjacent time points have higher correlation, while measurements at more distant time points have lower correlation. Ordinary linear regression models require that each observation be independent. This assumption clearly does not hold true for repeated measures data structures. Forcing the use of ordinary regression models for analysis can lead to biased estimation of parameter standard errors, thus affecting the accuracy of hypothesis testing and statistical inference. Therefore, this embodiment introduces a generalized estimating equation (GEE) model and sets an occupational correlation matrix to handle the correlation between repeated measures data, solving the modeling problem of intra-individual correlation in clinical longitudinal data. The specific implementation is as follows: Generalized Estimating Equations (GEEs) are statistical methods specifically developed for analyzing repeated measures and longitudinal data. Their core idea is to describe the correlation structure among multiple measurements of the same subject using a working correlation matrix (WCMM), and incorporate this structure into the model estimation, thereby estimating and inferring regression parameters while considering the inherent correlations within the data. The WCMM is a parameter matrix with dimensions equal to the number of repeated measures, where each element is the correlation coefficient between corresponding time points. A key characteristic of GEE is robustness: even if the configured WCMM does not perfectly match the true correlation structure, the regression parameter estimation remains consistent, only the estimation efficiency may decrease. Since the true correlation structure of clinical data is often unknown, researchers can only configure it based on experience. The robustness of GEE ensures that the analytical conclusions remain reliable even when the correlation matrix configuration has some bias, thus possessing high practical value.

[0044] Obtain repeated measures data from the same patient at different time points within the merged analysis dataset. Repeated measures data refers to the set of values ​​for the same set of clinical indicators recorded at different measurement time points within the dataset, all with the same patient identifier. For example, for a patient randomly assigned to the treatment group, measurements of various indicators such as weight, body mass index, luteinizing hormone, follicle-stimulating hormone, and testosterone were recorded at four time points: baseline, week 4, week 8, and week 12. These measurements at these four time points constitute the patient's repeated measures data.

[0045] A GEE model was constructed, with treatment modality and measurement time as independent variables and the changes in each clinical indicator as dependent variables. An operation-related matrix was also established among the repeated measures data. Including treatment modality in the model aims to assess the differences in the impact of different treatment regimens on indicator changes; this is typically coded as a dummy variable. Including measurement time aims to control for and isolate the natural influence of time trends on the indicators, eliminating confounding effects from changes in indicators caused by the passage of time itself. The operation-related matrix can be selected based on data characteristics and professional knowledge.

[0046] The effect estimates for each independent variable are calculated using the GEE model, and their absolute values ​​are used as the first importance score for each indicator. The magnitude of the effect estimate reflects the average magnitude of the change in the dependent variable when the independent variable changes by one unit; the larger the absolute value, the stronger the impact. In this embodiment, the effect estimate corresponding to the treatment method directly reflects the difference in the impact of different treatment plans on the degree of improvement of the indicator. The greater the difference, the stronger the indicator's ability to distinguish the efficacy of different treatment methods, and the more important it is in the evaluation system. This score will be passed to the subsequent normalization and fusion steps.

[0047] This embodiment effectively addresses the problem of biased parameter standardization caused by neglecting intra-cluster correlations in ordinary regression by constructing a GEE model to process repeated measures data. GEE introduces an occupational correlation matrix to model the correlations between repeated measures, ensuring that parameter estimation fully considers the hierarchical structure of the data, thus yielding more accurate and reliable effect estimates. The robustness of GEE guarantees consistent parameter estimation even with slight deviations in the occupational correlation matrix setting, reducing the technical threshold and risk of model setting. Simultaneously incorporating treatment methods and measurement times effectively separates treatment effects from time effects, allowing effect estimates to more accurately reflect the net impact of treatment itself on indicator changes. Using the absolute value of the effect estimate as the first importance score ensures that the score directly derives from statistically significant quantitative results, with intuitive clinical interpretation, providing a solid foundation for integration with machine learning results.

[0048] In some preferred embodiments, setting the job correlation matrix between the repeated measurement data includes: obtaining the measurement time interval of the repeated measurement data; if the measurement time intervals are not equidistant, then selecting an unstructured job correlation matrix; if the measurement time intervals are equidistant, then calculating the correlation coefficient between each measurement time point, and selecting an autoregressive job correlation matrix or a commutative job correlation matrix based on the changing trend of the correlation coefficient between each measurement time point with the measurement time interval.

[0049] In practice, while the specific form of the job-related matrix in the GEE model does not affect the consistency of parameter estimation, it significantly impacts estimation efficiency. If the chosen matrix form closely approximates the true correlation structure, parameter estimation efficiency is high, characterized by a small standard error, narrow confidence interval, and high test power. Conversely, if the form significantly deviates from the true structure, although the estimation remains consistent, the standard error increases, test power decreases, and the true treatment effect may be missed. This embodiment dynamically selects the matrix type based on the measurement time interval and correlation coefficient characteristics, maximizing estimation efficiency while ensuring consistency, and improving the accuracy and test power of the first importance score, as detailed below: The measurement time interval for obtaining repeated measures data. The measurement time interval refers to the length of time between two adjacent measurement time points in the dataset. For example, if the measurement time points in a clinical trial are set at baseline, week 4, week 8, and week 12, then the interval between all adjacent measurement time points is 4 weeks, which is considered an equidistant measurement time interval. If the time intervals for each follow-up are not entirely consistent due to limitations in actual trial execution conditions—for example, some patients' week 4 follow-up is actually completed at week 3, and their week 8 follow-up is actually completed at week 9—then the measurement time intervals are not equidistant.

[0050] If the measurement time intervals are not equal, an unstructured working correlation matrix is ​​selected. This matrix treats the correlation coefficient between each pair of time points as an independent parameter, directly estimated from the data, without imposing any preset constraints. When the intervals are not equal, the correlation does not have a fixed decay law, and the unstructured matrix can accurately reflect the true correlation between any pair of time points in the most flexible way. Its limitation is that there are many parameters to be estimated, which increase quadratically with the number of repeated measurements, which may lead to difficulties in computational convergence.

[0051] If the measurement time intervals are equidistant, the correlation coefficient between each measurement time point is calculated. Based on the trend of the correlation coefficient between each measurement time point with the measurement time interval, an autoregressive working correlation matrix or a commutative working correlation matrix is ​​selected. Specifically, the correlation coefficients of adjacent time points (first group) and the correlation coefficients of non-adjacent time points separated by one time point (second group) are calculated, and the decrease in the second group relative to the first group is compared. If the decrease exceeds a preset decay threshold (e.g., 20% or 30%), it indicates that the correlation significantly decreases with increasing interval, and an autoregressive working correlation matrix is ​​selected. This matrix assumes that the correlation between adjacent points is the strongest and decreases exponentially with increasing interval. If the decrease does not exceed the threshold, it indicates that the correlation coefficients between time points are roughly similar and there is no obvious decay trend, and a commutative working correlation matrix is ​​selected. This matrix assumes that the correlation between any two different time points is equal and does not consider the effect of interval. Both autoregressive and commutative matrices only require the estimation of a small number of parameters, and have higher estimation efficiency and computational stability compared to unstructured matrices.

[0052] This embodiment addresses the problem of low estimation efficiency caused by improper matrix selection in GEE by establishing a dynamic selection mechanism for the job correlation matrix based on measurement time intervals and correlation coefficient characteristics. When the intervals are unequal, the unstructured matrix adapts to arbitrarily complex correlation patterns with a highly flexible parameterization method, ensuring accurate description. When the intervals are equal, the system automatically determines whether to select an autoregressive or commutative structure by calculating the actual correlation coefficients and observing decay characteristics, avoiding the uncertainty of subjective judgment. The autoregressive matrix uses a concise exponential decay model to characterize the natural decay of longitudinal data correlation, while the commutative matrix describes a roughly constant correlation structure with minimal parameters; both trade model complexity for higher estimation efficiency. Matching the appropriate matrix type to the data characteristics enables GEE to achieve narrower confidence intervals and higher test power while maintaining consistency, thereby improving the accuracy and discriminative power of the first importance score and making subsequent integration with machine learning results more reliable.

[0053] In some preferred embodiments, the machine learning algorithm includes a random forest. Based on the machine learning algorithm, a second feature importance analysis is performed on the merged analysis dataset to obtain a second importance score for each indicator. This includes: extracting multiple bootstrap samples from the merged analysis dataset; constructing a classification and regression decision tree based on each bootstrap sample; and randomly selecting a subset of features from the feature set to select the optimal splitting feature at each node of the classification and regression decision tree, thus constructing a random forest model; and calculating the average reduction in impurity of each feature in all decision trees within the random forest model, which is used as the second importance score for each indicator.

[0054] In clinical data on obesity-related polycystic ovary syndrome (PCOS), the impact of various indicators on treatment efficacy may exhibit complex nonlinear patterns and interactions between indicators. For example, the contribution of weight loss to overall efficacy may not be a simple linear relationship, but rather exhibit a threshold effect, meaning that the marginal contribution to efficacy begins to decrease or increase after weight loss reaches a certain threshold. Similarly, there may be a synergistic effect between improvement in insulin resistance index and restoration of sex hormone levels; simultaneous improvement in both may contribute more to efficacy than the simple sum of their individual contributions. Traditional linear statistical models struggle to effectively describe and handle these complex data relationships. Therefore, this embodiment utilizes the advantages of ensemble learning by constructing a random forest model and calculating the reduction in average impurity to achieve a nonparametric evaluation of indicator importance, as detailed below: Random Forest is a machine learning algorithm based on ensemble learning, consisting of multiple Classification and Regression Decision Trees (CART). Ensemble learning achieves better overall performance than a single learner by combining the predictions of multiple base learners. Random Forest incorporates a dual randomization mechanism in its construction to enhance diversity and generalization ability: the first layer is sample randomization, using bootstrap sampling to draw training samples with replacement, where approximately 2 / 3 of the data is selected in each sampling, with the remaining approximately 1 / 3 being called out-of-bag data; the second layer is feature randomization, where a subset of features is randomly selected from all features when the decision tree splits at each node, and then the optimal splitting feature is chosen from that subset. Sample randomization ensures that each tree is trained on a different subset of data, while feature randomization introduces differences in splitting strategies. The combination of these two mechanisms reduces inter-tree correlation, decreases the variance of the ensemble model, and prevents overfitting of individual decision trees.

[0055] Multiple bootstrap samples are extracted from the merged analysis dataset, and a Classification and Regression Decision Tree (CART) is constructed based on each bootstrap sample. The construction of the decision tree is a top-down recursive splitting process: at each node, all features and their possible split points in the candidate feature subset are traversed, the impurity reduction of each splitting scheme is calculated, and the feature and split point with the largest reduction are selected for splitting, dividing the data into two child nodes. Impurity measures the consistency of the sample output values ​​within a node; mean squared error is commonly used for regression problems, and the Gini coefficient is commonly used for classification problems. During node splitting, a feature subset is randomly selected from the feature set to choose the optimal splitting feature, thus constructing the random forest model.

[0056] The average reduction in impurity for each feature across all decision trees in the random forest model is calculated and used as the second importance score for each indicator. For a decision tree, the reduction in impurity for a feature is equal to the sum of the reductions in impurity before and after the split at all nodes where that feature was selected as the splitting feature. The greater the reduction in impurity for a feature in a single split, the more effectively it can distinguish mixed samples, meaning the greater its contribution to the prediction results. The average reduction in impurity for the same feature across all decision trees is summed and divided by the total number of decision trees, and this average reduction in impurity is used as the second importance score for that feature. The second importance score reflects the predictive importance of each indicator to the therapeutic outcome from a data-driven ensemble learning perspective. It will then be passed to subsequent normalization and weighted fusion steps and fused with the first importance score.

[0057] This embodiment objectively evaluates the importance of indicators from an ensemble learning perspective through a random forest model and average impurity reduction. The dual randomization mechanism significantly reduces inter-tree correlation and ensemble model variance by randomly superimposing samples and features, effectively avoiding overfitting. Compared to traditional linear models, random forests do not require pre-defined variable relationships and distribution assumptions, and can automatically capture complex nonlinear and interaction effects through data-driven approaches, making them particularly suitable for obese polycystic ovary syndrome with complex pathological mechanisms and multiple intertwined factors. The average impurity reduction is based on the information gain of the model splitting process, offering high computational efficiency and intuitive results, quantifying the marginal contribution of each feature to the predictive performance of the ensemble model. This second importance score and the first importance score are independently evaluated from both data-driven and model inference perspectives, providing complementary information for subsequent fusion.

[0058] In some preferred embodiments, calculating the average reduction in impurity of each feature in the random forest model across all decision trees, as the second importance score for each indicator, includes: when constructing the random forest model, for each decision tree, calculating the prediction error of the decision tree using out-of-bag data that was not sampled; randomly shuffling the feature values ​​of a certain feature in the out-of-bag data, and recalculating the prediction error of the decision tree after shuffling; calculating the increase in the prediction error before and after shuffling, and calculating the average of the increase across all decision trees, using the average as the second importance score for each indicator.

[0059] To avoid overfitting bias caused by using only training data, this embodiment calculates the prediction error of each decision tree using out-of-bag data (not included in the training set) when constructing the random forest model. Out-of-bag data refers to approximately one-third of the original records from the bootstrap sampling that did not enter the training set. Since out-of-bag data has never participated in the tree's training and is completely independent, the prediction error calculated based on it accurately reflects the tree's generalization performance on independent new data and is unaffected by overfitting. For continuous dependent variables, the mean squared error (MSE) is commonly used to measure the prediction error; it is the mean of the squares of the differences between the true and predicted values.

[0060] Then, the feature values ​​of a specific feature in the out-of-bag data are randomly shuffled, and the prediction error of the decision tree is recalculated. During shuffling, other features and the dependent variable remain unchanged; only the target feature value is randomly rearranged among the samples. After rearrangement, any association between this feature and the dependent variable is destroyed, while the numerical distribution of the feature itself remains unchanged. If this feature is important for prediction, shuffling will lead to a significant decrease in predictive ability and a marked increase in error; if it is not important, the model can rely on other features to maintain accurate predictions, and the error will not change significantly.

[0061] The increase in prediction error before and after shuffling (error after shuffling minus error before shuffling) is calculated, and this increase is averaged across all decision trees to determine the second importance score for that feature. A larger positive value indicates a greater contribution of the feature to the tree's predictions. The average increase in prediction error is directly based on the model's performance change on independent validation data, and compared to the average reduction in impurity based on node splitting in the training data, it more accurately reflects the predictive value of the feature in real-world application scenarios.

[0062] This embodiment addresses the overfitting bias that can arise from importance assessments based on training node splitting by employing an evaluation method that utilizes changes in prediction error based on out-of-bag data. Out-of-bag data, naturally generated as an internal validation set during random forest construction, eliminates the need for additional data splitting or cross-validation, allowing for objective evaluation of generalization performance. This approach fully leverages all training data while providing an independent evaluation method. Based on the mechanism of error changes before and after shuffling feature values, the actual performance degradation of a single feature after its destruction is directly observed, quantifying its true contribution in a logically intuitive and clear manner. Since the evaluation is performed entirely on out-of-bag data not used in training, the results are unaffected by accidental fitting of the model to training noise, enabling more accurate identification of indicators with genuine predictive value on independent data and avoiding the misattribution of high scores to noise features that are only incidentally important on training data.

[0063] S3. Based on the optimized index weights, the weight values ​​in the original efficacy evaluation index system are replaced to generate the optimized efficacy evaluation index system.

[0064] In practice, the original efficacy evaluation index system refers to the multi-dimensional efficacy evaluation index system constructed based on retrospective clinical data through statistical screening and expert consultation. This system includes preset evaluation dimensions, the original weight values ​​of each indicator item, and the corresponding scoring rules. The replacement operation refers to assigning the optimized indicator weights obtained by the algorithm to each indicator item one by one, replacing the original weight values, and adjusting the scoring parameters and calculation formulas in the scoring rules accordingly based on the replaced weight values, so that the entire evaluation system is consistent with the new weight distribution, ultimately forming the optimized efficacy evaluation index system.

[0065] The method provided in this embodiment systematically integrates retrospective and prospective datasets, and uses multivariate statistics and machine learning algorithms to perform feature importance analysis and regression fitting on the merged data to reweight it. This solves the problem that existing indicator systems rely solely on retrospective data and lack prospective validation and objective calibration. Retrospective real-world data has a large sample size and broad coverage, while prospective RCT data has a high level of causal inference evidence and standardized collection. The fusion of the two provides both sample breadth and evidence strength, offering a more comprehensive and balanced data foundation for weight recalculation than a single data source. The combination of multivariate statistics and machine learning overcomes the limitations of a single algorithm: the former can handle repeated measures correlations and provide effect estimates with clear statistical meaning, while the latter can capture complex nonlinear relationships and higher-order interaction effects. The two evaluate and fuse the results from different perspectives, making the optimized weights more robust and reliable than a single method. The optimized indicator system retains the original framework and dimensional divisions, and incorporates correction information from higher-level evidence-based evidence through weight replacement, achieving dynamic evolution and continuous improvement in scientific applicability.

[0066] In some preferred embodiments, the step of replacing the weight values ​​in the original efficacy evaluation index system based on the optimized index weights to generate an optimized efficacy evaluation index system includes: obtaining the original efficacy evaluation index system, which includes preset evaluation dimensions, the original weight values ​​of each index item, and scoring rules; replacing the original weight values ​​with the optimized index weights; and updating the scoring rules based on the replaced index weights to generate the optimized efficacy evaluation index system.

[0067] After obtaining the optimized indicator weights through algorithmic recalculation, the next step is to systematically apply these new weight values ​​to update the initially established efficacy evaluation indicator system. This ensures that the new weight values ​​are accurately mapped to each indicator item in the original system, avoiding weight mismatch due to incorrect indicator item matching. Simultaneously, it ensures that the updated scoring rules remain completely consistent with the new weight values, enabling the entire evaluation system to be directly used clinically. To this end, this embodiment achieves smooth updates and dynamic evolution of the indicator system through a standardized update process: acquiring the original system, replacing weight values ​​item by item, and updating the scoring rules. The specific implementation method is as follows: The existing efficacy evaluation index system was obtained, which includes pre-set evaluation dimensions, the original weight values ​​of each indicator, and scoring rules. The pre-set evaluation dimensions include disease indicator dimensions, syndrome dimensions, symptom dimensions, and detection indicator dimensions. The disease indicator dimension covers core indicators directly related to the pathological mechanism of obesity-related polycystic ovary syndrome, including the insulin resistance index and glucose tolerance. The syndrome dimension covers syndrome assessment indicators corresponding to TCM syndrome differentiation, including severity scores for spleen deficiency with phlegm-dampness syndrome and kidney deficiency with blood stasis syndrome. The symptom dimension covers quantitative assessment indicators of patients' subjective symptoms and observable signs, including menstrual cycle recovery, hirsutism symptom scores, and acne severity grading. The detection indicator dimension covers objective physicochemical indicators obtained through laboratory testing, including luteinizing hormone (LH) levels, follicle-stimulating hormone (FSH) levels, testosterone levels, progesterone levels, blood lipid biochemical indicators, uric acid levels, and fatty liver severity. The original weight values ​​are relative importance coefficients of each indicator in the evaluation system, determined jointly by statistical screening of retrospective clinical data and expert consultation. Expert consultation can employ the Delphi method, a structured group decision-making method that gradually converges expert opinions through multiple rounds of anonymous questionnaire feedback. The scoring rules define a series of calculation methods and judgment criteria for calculating the scores of each indicator based on the patient's actual clinical data and summarizing them into a total efficacy evaluation score. These include the grading and scoring standards for each indicator and the formula for calculating the total efficacy evaluation score.

[0068] The original weight values ​​are replaced with the optimized indicator weights. The replacement operation is performed one-to-one according to the indicator items. Specifically, it iterates through all indicator items under all dimensions in the original evaluation indicator system. For each indicator item, it finds the corresponding new weight value in the optimized indicator weights and updates the original weight value of the indicator item with the found new weight value. For indicator items newly added in the optimized indicator weights but not included in the original system, the indicator item is added to the original system according to its evaluation dimension and assigned the corresponding weight value. For indicator items that exist in the original system but have been removed from the optimized indicator weights, the indicator item is removed from the system or its weight value is set to zero to preserve the integrity of the system structure.

[0069] Based on the replaced indicator weights, the scoring rules are updated to generate an optimized efficacy evaluation indicator system. The specific steps of updating the scoring rules include recalculating the full score allocation ratio for each evaluation dimension according to the new weight values, re-determining the highest assigned score for each indicator item, adjusting the score interval divisions in the grading scoring criteria for each indicator item, and updating the weighting coefficients in the efficacy evaluation total score calculation formula according to the new weight values. After the update, all scoring parameters and calculation coefficients in the scoring rules are consistent with the replaced indicator weights, and the entire indicator system constitutes a self-consistent and directly applicable complete evaluation framework for clinical scoring.

[0070] This embodiment achieves systematic and dynamic optimization of the indicator system through a standardized process of acquiring the original system, replacing weights item by item, and synchronously updating the scoring rules. The item-by-item replacement mechanism ensures accurate mapping between optimized weights and original indicator items, avoiding mismatches or omissions. Synchronous updating of the scoring rules guarantees consistency between the scoring criteria and the new weight distribution, ensuring the evaluation system remains logically self-consistent after the update. The entire update process is clear and does not involve restructuring the evaluation framework or dimensional divisions, maintaining structural continuity and the continuity of usage methods, thus reducing the clinical adaptation costs associated with changes. This update mechanism has excellent repeatability: when new data is accumulated in the future or a larger-sample prospective trial is completed, the complete process of data merging, algorithm recalculation, and weight replacement can be repeatedly executed to perform a new round of updates and calibrations to the indicator system, allowing it to continuously evolve and iterate with the accumulation of evidence-based data, always maintaining its optimal state supported by the best current evidence.

[0071] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0072] Please see Figure 2 , Figure 2This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0073] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0074] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it causes the processor 502 to execute an optimization method for a therapeutic efficacy evaluation index system.

[0075] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0076] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute an optimization method for an efficacy evaluation index system.

[0077] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0078] The processor 502 is used to run a computer program 5032 stored in a memory to implement the steps of the method for optimizing an efficacy evaluation index system provided in any of the above method embodiments.

[0079] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0080] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0081] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the steps of the optimization method for an efficacy evaluation index system provided in any of the above method embodiments.

[0082] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0083] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0084] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0085] The method steps in the embodiments of the present invention can be adjusted, merged, or reduced in order according to actual needs. The device units in the embodiments of the present invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0086] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0087] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, this invention is also intended to include these modifications and variations as long as they fall within the scope of the claims of this invention and their equivalents.

[0089] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for optimizing a therapeutic efficacy evaluation index system, characterized in that, include: Obtain the merged analysis dataset, which is obtained by merging retrospective clinical datasets and prospective randomized controlled trial datasets. The merged analysis dataset contains clinical indicator data related to multiple evaluation dimensions recorded under multiple treatment modalities. Based on multivariate statistical methods and machine learning algorithms, feature importance analysis and regression fitting are performed on the merged analysis dataset. The influence coefficients and correlation weights of each indicator on the efficacy results are recalculated to obtain the optimized indicator weights. Based on the optimized index weights, the weight values ​​in the original efficacy evaluation index system are replaced to generate the optimized efficacy evaluation index system.

2. The method according to claim 1, characterized in that, The retrospective clinical dataset and the prospective randomized controlled trial dataset are merged to obtain the merged analysis dataset, which includes: Obtain the retrospective clinical dataset and the prospective randomized controlled trial dataset; Field mapping is performed on the retrospective clinical dataset and the prospective randomized controlled trial dataset to uniformly map field names with the same clinical meaning to preset standardized fields; The mapped data is normalized to eliminate the dimensional differences between the retrospective clinical dataset and the prospective randomized controlled trial dataset; The normalized data is cleaned and verified to obtain the merged analysis dataset.

3. The method according to claim 1, characterized in that, The merged analysis dataset is subjected to feature importance analysis and regression fitting based on multivariate statistical methods and machine learning algorithms. The influence coefficients and correlation weights of each indicator on the efficacy results are recalculated to obtain optimized indicator weights, including: The first feature importance analysis was performed on the merged analysis dataset based on multivariate statistical methods to obtain the first importance score of each indicator; The merged analysis dataset is subjected to a second feature importance analysis based on machine learning algorithms to obtain the second importance score of each indicator; The first importance score and the second importance score are normalized respectively to obtain the normalized first importance score and the normalized second importance score; The weighted sum of the normalized first importance score and the normalized second importance score is used to obtain the fused importance score of each indicator. The fusion importance score is used as the influence coefficient and correlation weight of each indicator on the therapeutic effect, and is used as the optimized indicator weight.

4. The method according to claim 3, characterized in that, The multivariate statistical method includes a generalized estimation equation. The first feature importance analysis is performed on the merged analysis dataset based on the multivariate statistical method to obtain the first importance score for each indicator, including: Obtain repeated measurement data of the same patient at different time points in the merged analysis dataset; A generalized estimation equation model was constructed, with treatment method and measurement time as independent variables and the changes in each clinical indicator as dependent variables, and the job correlation matrix between the repeated measurement data was set. The effect estimates of each independent variable are calculated using the generalized estimation equation model, and the absolute values ​​of the effect estimates are used as the first importance scores for each indicator.

5. The method according to claim 4, characterized in that, Setting the job correlation matrix between the repeated measurement data includes: The measurement time interval for obtaining the repeated measurement data; If the measurement time intervals are not equal, then select the unstructured operation correlation matrix; If the measurement time intervals are equal, the correlation coefficient between each measurement time point is calculated, and an autoregressive job correlation matrix or a commutative job correlation matrix is ​​selected based on the changing trend of the correlation coefficient between each measurement time point with the measurement time interval.

6. The method according to claim 3, characterized in that, The machine learning algorithm includes random forest. The second feature importance analysis is performed on the merged analysis dataset using the machine learning algorithm to obtain the second importance score for each indicator, including: Multiple bootstrap samples are extracted from the merged analysis dataset. A classification and regression decision tree is constructed based on each bootstrap sample. When splitting at each node of the classification and regression decision tree, a subset of features is randomly selected from the feature set to select the optimal splitting feature, thus constructing a random forest model. The average reduction in impurity of each feature in the random forest model across all decision trees is calculated and used as the second importance score for each indicator.

7. The method according to claim 6, characterized in that, The calculation of the average reduction in impurity of each feature in the random forest model across all decision trees, as the second importance score for each indicator, includes: When constructing the random forest model, for each decision tree, the prediction error of the decision tree is calculated using the out-of-bag data that was not selected in the sampling. The feature values ​​of a certain feature in the out-of-bag data are randomly shuffled, and the prediction error of the decision tree after shuffling is calculated again. Calculate the increase in prediction error before and after shuffling, and calculate the average of the increase across all decision trees. Use the average as the second importance score for each indicator.

8. The method according to claim 1, characterized in that, The step of replacing the weight values ​​in the original efficacy evaluation index system based on the optimized index weights to generate the optimized efficacy evaluation index system includes: Obtain the original efficacy evaluation index system, which includes preset evaluation dimensions, the original weight values ​​of each index item, and scoring rules; Replace the original weight values ​​with the optimized index weights; Based on the replaced indicator weights, the scoring rules are updated to generate an optimized efficacy evaluation indicator system.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-8.

Images